JP6480625B1 - Preprocessor for abnormal sign diagnosis system - Google Patents

Abstract

A user interface for visualizing a physical model for predicting abnormality is provided. In a preprocessor to which time-series sensor data is input, a latent variable generation unit that generates a latent variable by applying a non-linear function to the input sensor data, and the latent variable generation to the input sensor data. The latent variable generated in the section is added to form a variable, and a latent variable adding section that outputs a plurality of variables, and clustering that divides the plurality of variables output by the latent variable adding section into subsystems according to the correlation A processing unit, a relational expression estimation processing unit that generates a relational expression of a plurality of variables for each subsystem divided by the clustering processing unit, and a physical model, for each subsystem divided by the clustering processing unit, Display sensor data nodes and latent variable nodes in different shapes, and display correlations between variables with lines between nodes. Comprising a Gosuru physical model display unit. [Selection] Figure 1

Description

  The present invention relates to a preprocessor for an abnormal sign diagnosis system.

  In order to improve the operation rate and the production rate of the apparatus, an abnormality sign diagnosis technique is used as one of the techniques for improving the maintenance of the apparatus using a data analysis method. In the abnormal sign diagnosis technique, it is considered effective to use a physical model that reflects the configuration of the apparatus and the mechanism of operation in order to improve the accuracy of diagnosis.

  Here, the physical model is composed of variables such as the output, input, and sensor data of the device, and mathematical formulas that reflect the mechanism and state of the device using these variables and coefficients. In constructing such a physical model, it is difficult to construct a physical model that can be applied from the beginning, and the constructed physical model is often displayed, modified by the user, and reconstructed.

  On the other hand, applying a physical model for diagnosis is performed, and Patent Documents 1 and 2 describe that a physical model is applied to diagnosis and the result of diagnosis is displayed. .

JP2013-008098A JP 2010-186310 A

  As described in Patent Documents 1 and 2, it is possible to apply a physical model to diagnosis. However, even if the applied physical model is displayed as text, for example, it is difficult for the user to understand the state of the physical model and to correct it. In Patent Documents 1 and 2, there is no sufficient description regarding visualization of the physical model in consideration of the ease of correction.

  Therefore, an object of the present invention is to provide a user interface for visualizing a physical model for abnormal sign diagnosis.

  A pre-processor of a typical abnormality sign diagnosis system according to the present invention is a pre-processor to which time-series sensor data is input, a latent variable generation unit that generates a latent variable by applying a nonlinear function to the input sensor data; Between the input sensor data, the latent variable generated by the latent variable generating unit is added as a variable, and a plurality of variables are output between the latent variable adding unit that outputs a plurality of variables and the latent variable adding unit. As a physical model, a clustering processing unit that divides the data into subsystems according to correlation, a relational expression estimation processing unit that generates a linear relational expression of a plurality of variables for each subsystem divided by the clustering processing unit, Different shapes of sensor data nodes and latent variable nodes for each subsystem divided by the clustering processing unit Displaying, and characterized in that and a physical model display unit for controlling to display the correlation line between nodes between variables.

  According to the present invention, it is possible to provide a user interface for visualizing a physical model for abnormal sign diagnosis.

It is a figure which shows the structural example of an abnormality sign diagnostic system. It is a figure which shows the example of multidimensional time series sensor data. It is a figure which shows the example of a latent variable. It is a figure which shows the example of the relationship between sensor ID, a nonlinear function, and a latent variable. It is a figure which shows the example of the result of dividing | segmenting a variable into a cluster. It is a figure which shows the example of clustering of a variable. It is a figure which shows the example of a user interface. It is a figure which shows the example of visualization of a physical model. It is a figure which shows the example which displays a relational expression. It is a figure which shows the example of the screen of multidimensional time series sensor data input. It is a figure which shows the example of the screen of the relevant information input of a multidimensional sensor. It is a figure which shows the example of visualization of the physical model when a device structure is input. It is a figure which shows the example of the setting screen of the parameter for physical model construction. It is a figure which shows the example of the correction screen of multidimensional time series sensor data. It is a figure which shows the example of the correction screen of a latent variable. It is a figure which shows the example of the correction screen of a correlation coefficient. It is a figure which shows the example of the correction screen of a relational expression.

  The embodiment of the present invention is, for example, an abnormality sign diagnosis apparatus or an abnormality sign diagnosis system including a preprocessor and an abnormality sign diagnosis processing unit. An example of an embodiment of the preprocessor and the abnormal sign diagnosis processing unit may be a general computer, which may be a software implementation that the processor processes according to a program, or a dedicated hardware instead of a general computer May be implemented.

  Hereinafter, an example of an embodiment of the abnormality sign diagnosis processing system will be described in detail as an example with reference to the drawings.

  FIG. 1 is a diagram illustrating a configuration example of an abnormality sign diagnosis system. The input of the abnormality sign diagnosis system is multidimensional time series sensor data collected from the monitoring target device. For this purpose, the multidimensional sensor 3 is connected to the preprocessor 1, and the multidimensional time series sensor data output from the multidimensional sensor 3 is processed by the preprocessor 1, and the processing result is transferred from the preprocessor 1 to the abnormal sign diagnosis processing unit 2. Sent.

  In the preprocessor 1, first, the preprocessing unit 11 performs processing such as defect data processing, outlier removal, or normalization on multidimensional time series sensor data. Next, the latent variable adding unit 12 adds a new variable generated by the combination of nonlinear functions in the latent variable generating unit 19 to the multidimensional time series sensor data processed by the preprocessing unit 11. This new variable is called a latent variable. The latent variable will be further described later.

  The latent variable selecting unit 13 calculates a correlation coefficient of an arbitrary data set (data pair) among the data in which the latent variable is added to the multidimensional time series sensor data by the latent variable adding unit 12, and sets the set (pair). Are selected one by one, and information on the data to be set and the correlation coefficient are output to the clustering processing unit 14.

  Hereinafter, multidimensional time-series sensor data and latent variables included in the selected and output set are referred to as variables without being distinguished. The latent variable selection unit 13 does not need to select a set having a correlation coefficient lower than a preset threshold and does not output the set to the clustering processing unit 14.

  The clustering processing unit 14 processes information input from the latent variable selection unit 13 using a hierarchical clustering method, and divides a plurality of variables into a plurality of clusters based on correlation between variables. In this division, a plurality of related variables are included in one cluster, and a plurality of such clusters are generated.

  The relational expression estimation processing unit 15 estimates a relational expression between variables for each cluster divided by the clustering processing unit 14 using a linear regression analysis method. Thus, a physical model is generated. The physical model here includes a variable, a correlation coefficient between the variables, and a relational expression between the variables.

  The physical model display unit 16 displays the generated physical model as a part of a user interface (GUI) on a display device (not shown) or the like. The user interface here includes a user interface for modifying each part of the physical model.

  The user correction processing unit 17 receives a correction input to the physical model from the user via the user interface, analyzes the correction input of the user, generates a construction condition of the physical model from the analysis result of the correction input, and generates the generated building condition As a feedback to the latent variable adding unit 12.

  On the other hand, the latent variable adding unit 12 may transfer to the latent variable selecting unit 13, the clustering processing unit 14, and the relational expression estimating processing unit 15 according to the generated configuration condition, or the user correction processing. The unit 17 may transmit the generated configuration condition to the latent variable selection unit 13, the clustering processing unit 14, and the relational expression estimation processing unit 15 in addition to the latent variable adding unit 12.

  The display to the user and the input from the user are performed by another device, and the device and the preprocessor may communicate via a network, and the physical model display unit 16 and the user correction processing unit 17 display via the network. And the input may be controlled.

  When the multidimensional time series sensor data output from the multidimensional sensor 3 includes physical unit information of each multidimensional sensor 3, the physical unit extraction unit 18 extracts the physical unit of the multidimensional time series sensor data and generates a latent variable. To the unit 19. The latent variable generation unit 19 generates a latent variable according to a latent variable calculation procedure for a combination of physical units having a predefined physical meaning.

  FIG. 2 is a diagram illustrating an example of multidimensional time-series sensor data. The multi-dimensional time series sensor data includes time information 21 when each data (measurement value) is recorded, and measurement values 22 to 25 of each multi-dimensional sensor 3 at each recording time. As shown in FIG. 2, multidimensional data of the measurement values 22 to 25 is time-series data according to the time information 21.

  The multidimensional time series sensor data may include information related to the multidimensional sensor 3. For example, the physical unit (“bar”, “K”, etc.) of each multidimensional sensor 3 may be included as related information.

  FIG. 3 is a diagram illustrating an example of a latent variable. The time information 31 and the multidimensional time series sensor data 32 to 34 shown in FIG. 3 correspond to the time information 21 and the measured values 22 to 24 shown in FIG. The sensor ID of the sensor 3 is “X”, and “Y” and “Z” are also sensor IDs thereafter.

  The latent variable is generated by combining a value of multidimensional time series sensor data and a nonlinear function. For example, “F1 (X, Y)” and “F2 (Y, Z)” of the latent variables 35 to 38 shown in FIG. 3 are latent variables. Here, “F1” and “F2” represent nonlinear functions.

  For this reason, the value of the latent variable can be calculated from the value of the multidimensional time series sensor data for each recording time point. For example, when the time information 21 is “2017/01/01 08:00:00”, the multi-dimensional time-series sensor data 32 with the sensor ID “X” is “2.3”, and the sensor ID is “Y”. The multidimensional time series sensor data 33 is “65”.

  When these “2.3” and “65” are applied to the nonlinear function “F1”, the calculated value is “149.5”, which is the value of the latent variable 35. Since such a nonlinear function is applied, the value of the multidimensional time series sensor data and the value of the latent variable are time series values (data) in the same format.

  FIG. 4 is a diagram illustrating an example of the relationship among the sensor ID, the nonlinear function, and the latent variable. The sensor ID 41 shown in FIG. 4 corresponds to the sensor IDs of the multidimensional time-series sensor data 32-34 shown in FIG. 3, and the latent variable 44 shown in FIG. 4 is the latent variables 35-38 shown in FIG. This corresponds to the relationship between the nonlinear function and the sensor ID.

  The function ID 42 is an ID for identifying a nonlinear function, and the nonlinear function 43 is a calculation formula of the nonlinear function corresponding to each function ID. Here, “V1” or the like of the nonlinear function 43 is a calculation variable (function argument), “V1” is the first argument, and “V2” is the second argument.

  The calculation content of the nonlinear function whose function ID is “F1” is “V1 × V2” as shown in the nonlinear function 43, which is to multiply the value of the first argument by the value of the second argument. Means. Therefore, “F1 (X, Y)” of the latent variable 44 is a value of multidimensional time-series sensor data whose sensor ID 41 is “X” and a value of multidimensional time-series sensor data whose sensor ID 41 is “Y”. Is calculated as the multiplied value.

  The latent variable 44 “F4 (X, Y, Z)” is a value of multidimensional time-series sensor data with a sensor ID 41 of “X” and a value of multidimensional time-series sensor data with a sensor ID 41 of “Y”. And the value of the logarithm (“LOG”) of the value obtained by multiplying the value of the multi-dimensional time-series sensor data with the sensor ID 41 “Z”.

  The information of the sensor ID 41, the information of the latent variable 44, and the information of the function ID 42 and the nonlinear function 43 are set in advance, and the information of the set latent variable 44 is the multi-dimensional time of the sensor ID included in the information of the latent variable 44. Applying the value of the series sensor data, the value of the latent variable is calculated like the values of the latent variables 35 to 38 shown in FIG.

  FIG. 5 is a diagram illustrating an example of a result of dividing a variable into clusters. The table shown in FIG. 5 includes a subsystem ID 51 indicating the ID of the subsystem and a member 52 indicating the member included in the subsystem. Here, the subsystem is a collection of variables obtained by the cluster, and the subsystem corresponds to the cluster.

  In the example illustrated in FIG. 5, the subsystem having the subsystem ID 51 of “1” has four members “X”, “Z”, “F1 (X, Y)”, and “F2 (X, Y)” as members. It represents the inclusion of two variables, and there is a correlation between these four variables. Next, an example in which the subsystem ID 51 is “4” and “5” will be described with reference to FIG.

  FIG. 6 is a diagram illustrating an example of variable clustering. The plurality of variables are clustered according to the correlation coefficient and divided into clusters according to the correlation coefficient threshold. The correlation matrix 61 is a matrix representing a correlation coefficient between four variables “A” to “D”.

  First, for each correlation coefficient in the correlation matrix 61, a degree of difference of “difference = 1− | correlation coefficient |” is calculated. Next, among the sets (variable pairs) composed of “A” to “D”, the set having the lowest difference is extracted, and a link 65 is provided between “A” and “C”. Is set.

  Then, the set having the lowest difference is extracted from the set composed of the links 65, “B”, and “D”, and the link 66 is set between the links 65 and “B”. Is done. Here, the degree of difference between the link 65 and “B” is calculated by the minimum value of the degree of difference between “B” and “A” and “C”.

  Further, a set composed of the link 66 and “D” is extracted, and a link 67 is set between the link 66 and “D”. In order to divide a plurality of variables, a threshold value 68 for determining whether or not there is a correlation is set, a link 67 whose dissimilarity exceeds the threshold value 68 is deleted, and a variable in which the remaining links 65 to 66 are set is It is divided so that it becomes the same cluster.

  In the example illustrated in FIG. 6, division is performed such that “A”, “B”, and “C” are “cluster 4” and “D” is “cluster 5”. Since the cluster here reflects the subsystem of the device configuration, it will be called a subsystem from now on.

  Therefore, as shown in FIG. 5, the members of the subsystem whose subsystem ID 51 is “4” are the members 52 “A”, “B”, and “C”, and the subsystem ID 51 that is “5”. A member of the system becomes “D” of member 52.

  The relational expression estimation processing unit 15 estimates relational expressions between variables in each subsystem. For example, in the example shown in FIG. 6, in order to estimate the relational expression in “subsystem 4”, “A” is first set as an objective variable, and “B” and “C” are set as explanatory variables. . Then, an expression expressing “A” with “B” and “C” is obtained by the linear regression method.

  An example of the result of linear regression is “A = 0.2 × B + 0.4 × C−0.6”. Next, “B” is set as an objective variable, an expression expressing “B” with “A” and “C” is obtained, and an expression expressing “C” is also obtained. Finally, relational expressions having respective variables as objective variables are obtained for “subsystem 4”.

  The prediction accuracy of the obtained relational expression may be evaluated. For example, the prediction accuracy may be evaluated using the residual of linear regression as an index. Information on the obtained relational expression may be sent to the abnormality sign diagnosis processing unit 2 and used as a reference for a normal state in the abnormality sign diagnosis.

  FIG. 7 is a diagram illustrating an example of a user interface. The window 710 displays information related to “physical model construction” and receives an operation, and includes a model display area 711, an information display area 712, and three panels of “input”, “control”, and “correction”. Each panel includes a plurality of buttons.

  When the “end” button 713 is clicked, the processed result is saved, and the display of the window 710 is cleared from the screen. The physical model display unit 16 displays the visualization result of the constructed physical model (variable, correlation coefficient between variables, subsystem, relational expression between variables) in the model display area 711. An example of visualization will be described later with reference to FIGS.

  The physical model display unit 16 displays information (such as text variable names and expressions) before visualization of the physical model in the information display area 712, or a part of the display in the model display area 711 is selected with a mouse or the like. Then, detailed information corresponding to the selected portion is displayed in the information display area 712. Further, the information display area 712 may display process log information, error information, and the like.

  The user correction processing unit 17 performs the following processing when buttons included in the three panels are clicked. Specific processing contents will be described later with reference to FIGS. When the “sensor data” button 71 in the “input” panel is clicked, a sensor data reading screen is displayed as shown in FIG. When the “related information” button 72 is clicked, a screen for inputting related information of the multidimensional sensor 3 as shown in FIG. 11 is displayed.

  When the “model construction” button 73 in the “control” panel is clicked, a physical model is generated using the input multidimensional time series sensor data and related information. The generation of the physical model may be started by the preprocessing unit 11 and the physical unit extraction unit 18, and subsequent processing of each unit may be continued with respect to the start of these processings.

  When the “apply model” button 74 is clicked, the constructed physical model is transmitted to the abnormality sign diagnosis processing unit 2, and the physical model is applied to the abnormality sign diagnosis. When the “parameter setting” button 75 is clicked, a parameter for constructing a physical model can be confirmed and displayed as shown in FIG.

  When the “Sensor” button 76 in the “Correction” panel is clicked, a screen for selecting, correcting, and adding multidimensional time-series sensor data is displayed as shown in FIG. When the “latent variable” button 77 is clicked, data of the latent variable is displayed as shown in FIG. 15, and a screen for correcting, deleting, and adding the latent variable is displayed.

  When the “correlation coefficient” button 78 is clicked, the correlation display of the physical model is selected with the mouse as shown in FIG. 16, and a screen for correcting the selected correlation coefficient and the like is displayed. When the “relational expression” button 79 is clicked, the value of the relational expression of the physical model is selected with the mouse as shown in FIG. 17, and a screen for correcting the relational expression corresponding to the selected value is displayed. The

  FIG. 8 is a diagram illustrating an example of visualization of a physical model displayed in the model display area 711. Since FIG. 8 is a black and white drawing, in order to represent colors, as shown in the legend 81, the colors are distinguished by the fill pattern. The legend 81 is for explanation and is not displayed in the model display area 711.

  The example of FIG. 8 is a display of three subsystems, and each subsystem is displayed in a different color. The range of the subsystem is displayed by a square line with the color of the subsystem, and the subsystem number (ID) is shown outside the square line. In the square line, variables belonging to the subsystem, correlations between variables, and relational expressions between variables are displayed.

  Variables are displayed as nodes. The type of variable is distinguished by the outline of the node. For example, a node whose outer shape is a circle represents a sensor, and a node whose outer shape is an octagon represents a latent variable. The (outer) size of the node represents the importance of the variable. The importance here is an index defined when a physical model is constructed.

  For example, if the user sets the importance of each variable from 1 (reference) to 10 (very important), the size of the node is set by this importance value. In addition to the user setting, the importance level may be calculated using another index.

  The color of the node is the same as the color of the subsystem, but the darkness of the color of the node represents the estimation accuracy of the relational expression in which the variable represented by the node is the objective variable. The node displays the name of the variable. In the example illustrated in FIG. 8, the latent variable is displayed as “F1”, but may be displayed as “V1 × V2”.

  When a node is selected with the mouse, information related to the selected node is displayed in the information display area 712. The correlation is displayed as a line connecting the nodes. The width of the line represents the strength of the correlation and is set by the absolute value of the correlation coefficient. The darkness of the line represents the stability of the correlation. Stability is calculated from the time variation of the correlation coefficient.

  For example, a plurality of correlation coefficient values may be calculated within a preset period, and the line color density may be set using the standard deviation of the calculation result as an index. If there is information on the causal relationship between variables, the causal relationship may be displayed with an arrow attached to the line.

  FIG. 9 is a diagram illustrating a display example of a relational expression. In order to simplify the description, an example of display for only one subsystem is shown. When a node is clicked with the mouse, a relational expression having the clicked node as an objective variable is selected. For example, when a node having a variable “Z” is clicked, a relational expression having “Z” as an objective variable is selected.

  The outline of the node of the clicked objective variable (“Z”) is highlighted by a thick double line, but the related explanatory variables (“F1 (X, Y)” and “F2 (X, Y)”) The outline of the node is highlighted with a normal double line.

  The coefficients (“0.75” and “−0.3”) for each variable in the relational expression are displayed between the explanatory variable node and the target variable node. The constant of the relational expression (“0.6”) is displayed near the node of the objective variable. In the information display area 712, the selected relational expression is displayed as text.

  FIG. 10 is a diagram illustrating an example of a multi-dimensional time series sensor data input screen. A multi-dimensional time series sensor data input screen is displayed as a window 101. For example, when the “sensor data” button 71 is clicked in the window 710 shown in FIG. 7, the window 101 may be opened.

  Multidimensional time-series sensor data input from the multidimensional sensor 3 is stored in a file by the preprocessor 1. When the file is designated by a user operation using the “reference” button 103, the file name of the designated file is displayed in the column 102, and the data of the designated file is read and displayed in the column 104. The

  The “reference” button 103 may not be used, and a file name may be input as a character string in the field 102. The display content of the column 104 is the data shown in FIG. 2, but in the example of FIG. 10, each multidimensional sensor 3 is represented as “variable 1” or the like. When the “Clear” button 105 is clicked, the read data is deleted.

  When the “Save” button 106 is clicked, the read data is saved in a specified file as multidimensional time-series sensor data to be analyzed. The field 102 and the “reference” button 103 are used to specify the file to be saved. When the “Exit” button 107 is clicked, the window 101 is closed.

  FIG. 11 is a diagram illustrating an example of a related information input screen of the multidimensional sensor 3. In the window 111, related information 112, an apparatus configuration diagram input area 114, and a menu 113 are displayed. In the related information 112, information of a plurality of multidimensional sensors 3 is displayed in a table format.

  In the column of the table of the related information 112, information of each multidimensional sensor 3 is displayed. Information such as sensor name, display name, physical quantity, physical unit, and sampling rate is displayed in the table row. The contents of the table may be input or modified by a user operation.

  Items displayed on each multi-dimensional sensor 3, that is, rows in the table, may be added and deleted by using “+” and “−” buttons. The information on these items is preferably information on items used for physical model construction.

  The apparatus configuration diagram input area 114 is an area where the configuration of the apparatus to be diagnosed and the position of the multidimensional sensor 3 are input and displayed. A drawing of the device configuration (approximate) is input using “baht” of the menu 113, and the position of the multidimensional sensor 3 is input using “sensor” of the menu 113.

  For example, when the “part” figure in the menu 113 is selected with the mouse and the mouse is clicked at an arbitrary position in the device configuration diagram input area 114, the figure may be placed at the clicked position. The figure placed in the device configuration diagram input area 114 may be resized or moved.

  In the “parts” of the menu 113, figures of various shapes are listed, and the work placed in the device configuration diagram input area 114 is repeated to input a drawing close to the actual device configuration.

  The “sensor” in the menu 113 is a double circle as a figure representing the multidimensional sensor 3. The double circle in the “sensor” of the menu 113 is selected by the mouse, and the mouse is placed at an arbitrary position in the device configuration diagram input area 114. When is clicked, a double circle is placed at the clicked position.

  The position of the double circle may be moved, and a number is set for the double circle using a keyboard or the like. For example, when “1” is set as the number, it is associated with “variable 1” of the related information 112. Also good. This makes it possible to represent the multidimensional sensor 3 actually installed in the apparatus.

  Information on the shape of the figure, the size of the figure, the position of the figure, and the position of the multidimensional sensor 3 input in the device configuration diagram input area 114 may be set in advance. In particular, the position information of the multidimensional sensor 3 may be acquired together with information such as “physical quantity” of the multidimensional sensor 3 or may be included in the multidimensional time series sensor data.

  The preset information is displayed in the device configuration diagram input area 114 without being input. Although FIG. 8 shows an example of the physical model display based on the node and the correlation, it is also possible to visualize the physical model including the configuration of the apparatus and the position of the multidimensional sensor 3. An example of this display is shown below.

  FIG. 12 is a diagram illustrating an example of visualization of a physical model when a device configuration is input. FIG. 8 shows an example of visualization of “subsystem 1” and “subsystem 2”. If there is information on the configuration of the apparatus and the position of the multidimensional sensor 3 shown in FIG. 11, as shown in FIG. The physical model is visualized. The physical model shown in FIG. 12 is also displayed in the model display area 711.

  The display contents of the nodes representing the variables and the display contents of the lines representing the correlation are the same as in FIG. 8, but in the example shown in FIG. 12, the configuration of the device is displayed, and “subsystem 1” and “subsystem 2” are displayed. The variable node corresponding to the multidimensional sensor 3 is displayed at the position of the multidimensional sensor 3.

  It is preferable that the node of the latent variable is displayed in the middle of a plurality of nodes that are correlated and connected by a line, but is not limited to this, and a position that does not overlap with the display of other nodes is calculated. , May be displayed at the calculated position.

  FIG. 13 is a diagram illustrating an example of a screen for setting parameters for constructing a physical model. In the window 131, setting values regarding each parameter are displayed in a table format. The column of the table displayed in the window 131 includes a parameter name, a parameter value, and a setting column for whether to automatically adjust the parameter.

  The rows in the table are parameter types and settings for each parameter. When “automatic adjustment” of the parameter is “NO”, a value may be input to the “value” of the parameter with a keyboard or the like, or may be corrected. When the parameter “automatic adjustment” is “YES”, the parameter value is calculated and set.

  For example, the parameter “determining whether or not there is a correlation” may be the threshold value 68 shown in FIG. If “automatic adjustment” corresponding to “correlation presence / absence determination” is “YES”, the average value of 0 and 1 may be simply calculated and set, or the average value of the minimum value and the maximum value of the dissimilarity may be calculated The number of subsystems may be calculated and set so as to be a predetermined number.

  When the parameter “automatic adjustment” is “YES”, the value may be calculated so that the setting by the user, which will be described later with reference to FIG. 16, is reflected. When the “save” button 132 is clicked, the setting value displayed in the window 131 is saved, and the saved setting value is applied in the subsequent processing. Here, automatic adjustment information may be stored.

  When the “End” button 133 is clicked, the window 131 is closed. If the “end” button 133 is clicked without clicking the “save” button 132, the new set value is discarded without being saved even if the new set value is displayed. .

  FIG. 14 is a diagram illustrating an example of a screen for correcting multidimensional time-series sensor data. The window 141 displays the read multidimensional time series sensor data in a table format. That is, in the column 142 of the window 141, a table in which each multidimensional sensor 3, that is, the value of each variable and time is associated is displayed.

  In order to correct the multidimensional time series sensor data to be analyzed, the multidimensional sensor 3 may be added or deleted in the column 142.

  When the multidimensional sensor 3 is selected, that is, the column of the variable corresponding to the selected multidimensional sensor 3 is selected in the table of the column 142, and the “remove” button 144 is clicked, the selected multidimensional sensor 3 is selected. The multidimensional time-series sensor data of the three-dimensional sensor 3 is excluded from the analysis, and the display color of the characters is changed from black to gray.

  In addition, when a column of variables corresponding to the excluded multidimensional sensor 3 is selected and the “provide” button 143 is clicked, the multidimensional time-series sensor data of the excluded multidimensional sensor 3 is the target of analysis. Return.

  When a “+” button or an “add” button 145 is clicked and a sensor name or variable name of a new multidimensional sensor 3 is designated, a variable column corresponding to the designated multidimensional sensor 3 is displayed. Added.

  Further, when the “save” button 146 is clicked, the multidimensional time-series sensor data included in the column 142 that is not excluded in the column 142 is stored. When the “Exit” button 147 is clicked, the window 141 is closed.

  FIG. 15 is a diagram illustrating an example of a latent variable correction screen. The window 151 is the same as the window 141 shown in FIG. 14 except that a table in which the value of the latent variable is associated with the time is displayed in the column 152, and thus description thereof is omitted.

  FIG. 16 is a diagram illustrating an example of correcting the correlation coefficient on the visualization screen. The model display area 711 and the information display area 712 are as described above, and one physical model (subsystem) is visualized in the model display area 711.

  Here, for example, when a line (correlation) between “F1 (X, Y)” and “Z” of the physical model displayed in the model display area 711 is double-clicked, correction for the correlation is performed. Window 1601 is opened. Instead of double-clicking the line, the “correlation coefficient” button 78 in the “correction” panel of the window 710 shown in FIG. 7 may be clicked.

  The window 1601 displays two variables (here, “F1 (X, Y)” and “Z”), “correlation coefficient”, “correction value”, and “correlation presence / absence” that are correlated to be corrected. Including. For the display of “correlation coefficient”, the value of the correlation coefficient when the window 1601 is displayed is displayed.

  In the correction of the value of the correlation coefficient by inputting a numerical value from a keyboard or the like, a numerical value is input in a column for “corrected value” display. For the display of “correlation presence / absence”, the presence / absence of correlation at the time when the window 1601 is displayed is displayed. In the example shown in FIG. 16, “YES” is displayed to indicate that there is a correlation.

  When “automatic adjustment” in the window 131 shown in FIG. 13 is set to “YES”, the value for the “correction value” display is calculated regardless of the numerical value input for the “correction value” display. Is done.

  That is, if “automatic adjustment” in the window 131 is set to “YES” and the display of “correlation presence / absence” in the window 1601 is set to YES by a user operation, there is a correlation. If NO is set by the user's operation, a value with no correlation is calculated.

  For example, if the correlation coefficient of two variables is 0.7, but the threshold value for determining whether or not there is a correlation is set to 0.8, the two variables having a correlation coefficient of 0.7 are correlated. It is determined that there is not. Therefore, if “YES” is set by the user's operation with respect to the “correlation presence / absence” display, the correlation coefficient is too low. Therefore, the correlation coefficient exceeding 0.8 is calculated and set. The

  Note that the threshold for determining whether or not there is a correlation may be set to be less than 0.7 and adjusted. By such adjustment processing, the physical model can be corrected by trial and error and judgment based on experience even if the user has no knowledge of adjustment of coefficients and thresholds.

  When a “save” button 1602 is clicked, a value corresponding to the display of “corrected value” is saved and applied. When an “end” button 1603 is clicked, the window 1601 is closed.

  FIG. 17 is a diagram illustrating an example of a relational expression correction screen. The model display area 711 and the information display area 712 are as described above, and one physical model (subsystem) is visualized in the model display area 711 and displayed together with a coefficient for each variable of the relational expression and a constant for the relational expression. ing.

  For example, when the “relational expression” button 79 in the “correction” panel of the window 710 shown in FIG. 7 is clicked and a node, that is, a variable displayed in the model display area 711 is selected by the mouse, the selected variable is changed. A window 1701 for correcting a relational expression as an objective variable is opened.

  Window 1701 includes displays of “object variable”, “relational expression”, and “delete”. For the “object variable” display, the name of the selected variable is displayed. For the display of “relational expression”, a relational expression having the selected variable as a target variable is displayed, and the coefficient and constant of the displayed relational expression are displayed in a square column so that the value can be corrected.

  In addition, in order to delete the relational expression instead of correction, a setting of YES is accepted for the display of “delete”. When a “Save” button 1702 is clicked, the coefficients and constants of the relational expression are saved and applied. When an “end” button 1703 is clicked, the window 1701 is closed.

  As described above, even if multidimensional time-series sensor data alone is not sufficient for constructing a physical model, it is possible to construct a physical model by adding latent variables as variables. By displaying the variables of the physical model constructed in this way so that the multidimensional time series sensor data and the latent variables can be distinguished, the user can easily understand the physical model.

  In addition, displaying correlation information between variables makes it easier for the user to understand the physical model. Furthermore, by displaying the apparatus in which the multidimensional sensor is installed and the multidimensional sensor as a configuration diagram, the user can easily understand the abnormality sign diagnosis system.

  Even if a sufficient physical model could not be constructed at one time, addition and deletion of multidimensional time-series sensor data, addition and deletion of latent variables, and parameter setting, especially for the displayed physical model, The physical model can be corrected by setting the number of relations and setting the coefficients and constants of the relational expression.

  Further, the automatic parameter adjustment enables the physical model to be corrected without the user having knowledge of parameter adjustment.

DESCRIPTION OF SYMBOLS 1: Preprocessor, 11: Pre-processing part, 12: Latent variable addition part, 13: Latent variable selection part, 14: Clustering process part, 15: Relational expression estimation process part, 16: Physical model display part, 17: User correction process Part: 18: physical unit extraction part, 19: latent variable generation part

Claims (13)

In a preprocessor that receives time-series sensor data,
A latent variable generation unit that generates a latent variable by applying a nonlinear function to the input sensor data;
A latent variable adding unit that adds a latent variable generated by the latent variable generating unit to the input sensor data to obtain a variable, and outputs a plurality of variables;
A clustering processing unit that divides a plurality of variables output by the latent variable adding unit into subsystems according to a correlation coefficient ;
A relational expression estimation processing unit that generates a relational expression of a plurality of variables for each subsystem divided by the clustering processing unit;
As a physical model, for each subsystem divided by the clustering processing unit, the sensor data nodes and latent variable nodes are displayed in different shapes, and the correlation coefficient between the variables is displayed as a line between the nodes. And a physical model display unit. The preprocessor according to claim 1, wherein
The physical model display unit
A preprocessor that controls a plurality of subsystems to display different colors. The preprocessor according to claim 2, wherein
The physical model display unit
The node is displayed with a size according to the importance set in advance for the variable and a density according to the prediction accuracy based on the residual of the linear regression calculated to obtain the relational expression by the relational expression estimation processing unit. To control and
A preprocessor that controls to display a line between nodes with a width according to a value of a correlation coefficient between variables and a density according to a temporal change of the correlation coefficient between variables. The preprocessor according to claim 3, wherein
The physical model display unit
A preprocessor that controls to display a relational expression between variables for which correlation is displayed. The preprocessor according to claim 4, wherein
The physical model display unit
Controls the specified node to be highlighted and displayed,
Control to display a relational expression with the variable or latent variable corresponding to the specified node as the objective variable and the variable or latent variable corresponding to the non-specified node as the explanatory variable,
A preprocessor which controls to display coefficients and constants of a displayed relational expression together with lines and nodes between nodes. The preprocessor according to claim 3, wherein
The input sensor data includes sensor position information,
The physical model display unit
Control to display the configuration of the device in which the sensor is installed;
A preprocessor for controlling to display a node of sensor data based on position information of the sensor together with display of the configuration of the device. The preprocessor according to claim 3, wherein
Accepting operations for adding or deleting sensor information that is the source of sensor data,
Accepting an operation for inputting a graphic representing a device in which the sensor is installed, accepting an operation for inputting a position where the sensor added in the device is installed,
A preprocessor further comprising: a user correction processing unit for associating information on the added sensor with a position where the added sensor is installed. The preprocessor according to claim 7, wherein
The physical model display unit
Control to display the configuration of the device;
A preprocessor which controls to display a node of sensor data based on a position where the added sensor is installed together with display of a configuration of the device. The preprocessor according to claim 8, wherein
The user correction processing unit
A parameter including a threshold value for determining whether or not there is a correlation between a plurality of variables output by the latent variable adding unit, and a parameter value for determining a physical model displayed by the physical model display unit or A preprocessor characterized by receiving an adjustment instruction. The preprocessor according to claim 9, wherein
The user correction processing unit
An adjustment instruction is received, an instruction for specifying a correlation is received, and when an instruction that there is a correlation is received, a parameter that is determined to have a correlation is determined regardless of the value of the received parameter. preprocessor. The preprocessor according to claim 10, wherein
The user correction processing unit
A preprocessor that receives an operation of excluding specified sensor data from a processing target among a plurality of input sensor data, or an operation of returning specified sensor data to a processing target when excluded. The preprocessor according to claim 11, wherein
The user correction processing unit
Among the plurality of latent variables generated by the latent variable generation unit, accepting an operation for excluding the designated latent variable from the processing target, or accepting an operation for returning the designated latent variable to the processing target when excluded. Feature preprocessor. The preprocessor according to claim 12, wherein
The user correction processing unit
A preprocessor that accepts correction of coefficients and constants of a displayed relational expression in a relational expression that uses a variable corresponding to a specified node as a target variable.

Images