Factor analysis device and factor analysis method

Abstract

A factor analysis device (100) comprises: a data acquisition unit (110) that acquires a plurality of types of monitoring data pertaining to an object to be monitored; a variable setting unit (120) that generates, on the basis of the monitoring data, training-oriented monitoring data, factor analysis-oriented monitoring data, a training-oriented data set composed of explanatory variables and objective variables, and a factor analysis-oriented data set; a model training unit (130) that generates, on the basis of the training-oriented data set, a learning model that has been trained so as to receive an explanatory variable as input and output an objective variable; a contribution level calculation unit (150) that calculates a contribution level, which is the degree of influence an explanatory variable has on an objective variable, on the basis of causal relationship data indicating causal relationships for the plurality of types of monitoring data, the learning model, the training-oriented monitoring data, and the factor analysis-oriented monitoring data; a factor analysis unit (160) that calculates a factor level, which is the degree of being a factor, for each explanatory variable on the basis of the contribution level; and a factor analysis result output unit (170) that outputs a factor analysis result on the basis of the factor analysis-oriented data set, the contribution level, and the factor level.

Description

Factor analysis device and factor analysis method

[0001] The disclosed technology relates to a factor analysis technology.

[0002] In an event requiring problem solving, it is important to take appropriate measures to address the problem in the event in order to achieve a desired outcome. For example, in a plant or factory, if there is a need to improve the productivity or efficiency of the equipment, it is important to take appropriate measures or operations to address the factors that reduce productivity or efficiency. Thus, to solve the problem, it is necessary to identify the problem in the event. However, the event is often the result of multiple overlapping conditions, making it difficult to identify the problem. Therefore, factor analysis techniques have been known that use a learning model to quantitatively analyze the relationship between the event (objective variable) and several candidate problems (explanatory variables) that are thought to be related to it. Patent Document 1 describes a "defect factor analysis method and defect factor analysis device" that "extracts product quality variation factors" by "utilizing a machine-learned learning model." Specifically, the "Defective Factor Analysis Method and Defective Factor Analysis Apparatus" of Patent Document 1 "performs the following steps: acquiring a plurality of pieces of manufacturing condition data or a plurality of pieces of monitoring data obtained by monitoring the operation of a manufacturing facility; predicting the quality of a manufactured product by inputting the acquired plurality of pieces of manufacturing condition data or monitoring data into a learning model trained to output quality data indicating the quality of a product manufactured by the manufacturing facility when the plurality of pieces of manufacturing condition data or monitoring data are input; and calculating, using the learning model, the contribution of each of the plurality of pieces of manufacturing condition data or monitoring data to the quality data or anomaly score data output from the learning model" (Patent Document 1, Abstract, Solution). According to the "Defective Factor Analysis Method and Defective Factor Analysis Apparatus" of Patent Document 1, a contribution indicating the degree to which each factor (explanatory variable in the learning model) directly contributes to product quality (target variable in the learning model) is calculated.

[0003] Japanese Patent Application Laid-Open No. 2022-043848

[0004] Multiple factors may have interdependent relationships, such that one factor affects another. Conventional techniques, when multiple factors (explanatory variables) have interdependent relationships, do not take into account the indirect contributions based on the interdependent relationships, resulting in incorrect analysis results based on incorrect contributions. Such conventional techniques have the drawback of making it difficult to improve the accuracy of factor analysis.

[0005] The present disclosure is intended to solve the above-mentioned problems and aims to improve the accuracy of factor analysis.

[0006] The factor analysis device of the present disclosure includes: a data acquisition unit that acquires multiple types of monitoring data related to a monitoring target; a variable setting unit that generates, based on the monitoring data, training monitoring data, factor analysis monitoring data, a training dataset consisting of explanatory variables and dependent variables, and a factor analysis dataset; a model learning unit that generates, based on the training dataset, a learning model that is trained to input explanatory variables and output dependent variables; a contribution calculation unit that calculates a contribution degree, which is the degree of influence that an explanatory variable has on a dependent variable, based on causal relationship data indicating a causal relationship in the multiple types of monitoring data, the learning model, the training monitoring data, and the factor analysis monitoring data; a factor analysis unit that calculates a factor degree, which is the degree to which each explanatory variable is a factor, based on the contribution degree; and a factor analysis result output unit that outputs a factor analysis result based on the factor analysis dataset, the contribution degree, and the factor degree.

[0007] The present disclosure provides an effect of improving the accuracy of factor analysis.

[0008] FIG. 1 is a diagram illustrating an example configuration of a factor analysis apparatus 100 according to a first embodiment of the present disclosure. FIG. 2 is a diagram illustrating an example configuration of a factor analysis system 10 including a factor analysis apparatus 100A according to a first embodiment of the present disclosure. FIG. 3 is a flowchart illustrating an example of processing by the factor analysis apparatus 100, 100A according to the present disclosure. FIG. 4 is a diagram illustrating a first example of causal relationship data used in the factor analysis apparatus 100, 100A according to the present disclosure. FIG. 5 is a diagram illustrating a second example of causal relationship data used in the factor analysis apparatus 100, 100A according to the present disclosure. FIG. 6 is a diagram illustrating a third example of causal relationship data used in the factor analysis apparatus 100, 100A according to the present disclosure. FIG. 7 is a diagram illustrating contribution degrees output by the contribution degree calculation unit 150, 150A in the factor analysis apparatus 100, 100A according to the present disclosure. FIG. 8 is a diagram illustrating factor degrees output by the factor analysis unit 160, 160A in the factor analysis apparatus 100, 100A according to the present disclosure. FIG. 9 is a diagram illustrating a first display example of the factor analysis result output by the factor analysis result output unit 170, 170A in the factor analysis apparatus 100, 100A of the present disclosure. FIG. 10 is a diagram illustrating a second display example of the factor analysis result output by the factor analysis result output unit 170, 170A in the factor analysis apparatus 100, 100A of the present disclosure. FIG. 11 is a diagram illustrating a configuration example of a factor analysis system 10 including a factor analysis apparatus 100B according to a second embodiment of the present disclosure. FIG. 12 is a diagram illustrating a display example of the factor analysis result output by the factor analysis result output unit 170B in the factor analysis apparatus 100B according to the second embodiment of the present disclosure. FIG. 13 is a diagram illustrating a configuration example of a factor analysis system 10 including a factor analysis apparatus 100C according to a third embodiment of the present disclosure. FIG. 14 is a diagram illustrating aggregation conditions used in the factor analysis apparatus 100C according to the third embodiment of the present disclosure. FIG. 15 is a diagram illustrating a configuration example of a factor analysis system 10 including a factor analysis apparatus 100D according to a fourth embodiment of the present disclosure. 16 is a diagram illustrating a first display example of the factor analysis result output by the factor analysis result output unit 170B in the factor analysis apparatus 100D according to the fourth embodiment of the present disclosure. FIG. 17 is a diagram illustrating a second display example of the factor analysis result output by the factor analysis result output unit 170B in the factor analysis apparatus 100D according to the fourth embodiment of the present disclosure.Fig. 18 is a diagram illustrating a first example of a hardware configuration for realizing functions according to the configuration of the present disclosure. Fig. 19 is a diagram illustrating a second example of a hardware configuration for realizing functions according to the configuration of the present disclosure.

[0009] In order to explain the present disclosure in more detail, embodiments of the present disclosure will be described below with reference to the accompanying drawings.

[0010] The present disclosure describes a technique for estimating the influence of explanatory variables on a target variable with higher accuracy than conventional techniques, using causal relationship data such as a causal graph.

[0011] First Embodiment In a first embodiment, a basic configuration of a factor analysis device according to the present disclosure will be described.

[0012] 1 is a diagram illustrating an example configuration of a factor analysis apparatus 100 according to a first embodiment of the present disclosure. The factor analysis apparatus 100 quantifies the magnitude of the influence of the monitoring data on the target by using a causal graph (causal relationship data) relating to the causal relationships or dependencies between objective variables, which are each target value, and explanatory variables, which are the monitoring data.

[0013] The factor analysis device 100 includes a data acquisition unit 110, a variable setting unit 120, a model learning unit 130, a causal relationship data acquisition unit 140, a contribution calculation unit 150, a factor analysis unit 160, and a factor analysis result output unit 170.

[0014] The data acquisition unit 110 acquires multiple types of monitoring data relating to the monitoring target.

[0015] The variable setting unit 120 generates, based on the monitoring data, training monitoring data, factor analysis monitoring data, a training dataset, and a factor analysis dataset. The training dataset is a dataset consisting of explanatory variables and a target variable.

[0016] The model learning unit 130 generates a learning model that is trained to receive explanatory variables as input and output a target variable based on the learning dataset.

[0017] The causal relationship data acquisition unit 140 acquires causal relationship data that indicates causal relationships among multiple types of monitoring data.

[0018] The contribution calculation unit 150 calculates the contribution, which is the degree of influence that an explanatory variable has on a target variable, based on causal relationship data indicating causal relationships in multiple types of monitoring data, the learning model, the monitoring data for learning, and the monitoring data for factor analysis.

[0019] The factor analysis unit 160 calculates a factor degree, which is the degree to which each of the explanatory variables is a factor, based on the contribution degree.

[0020] The factor analysis result output unit 170 outputs the factor analysis result based on the factor analysis data set, the contribution degree, and the factor degree.

[0021] The factor analysis device 100 is used, for example, for general production equipment in a factory that consumes resources such as electricity, gas, and oil. Monitoring data is collected from sensors installed in the production equipment and by monitoring the production equipment. For example, the factor analysis device 100 collects monitoring data obtained by monitoring the production equipment to be analyzed, i.e., the production equipment whose resource consumption efficiency and productivity are to be improved. The monitoring data includes sensor data such as resource consumption, temperature, and humidity collected by sensors installed in the equipment, as well as production management information such as product quality, production quantity, work time, workers, and production conditions. From the monitoring data, one or more data representing some resource consumption efficiency or productivity is set as a dependent variable, and two or more data related to the dependent variable, i.e., considered as candidates for factors that cause the dependent variable to fluctuate, are set as explanatory variables. The factor analysis device 100 trains a learning model using the explanatory variables as inputs and the dependent variable as outputs. The factor analysis device 100 calculates the degree of influence that explanatory variables have on the target variable, i.e., the "contribution of factors that fluctuate resource consumption efficiency and productivity," for each data sample and explanatory variable based on the learning model, monitoring data, and a causal graph between the monitoring data. The factor analysis device 100 aggregates the contributions and calculates the factor level for each explanatory variable. The factor analysis device 100 estimates the factors that fluctuate resource consumption efficiency and productivity, and presents information about the analysis results to users such as operators, such as on-site maintenance personnel for the target equipment, and resource managers. In this way, the factor analysis device 100 supports operators in their work to improve resource consumption efficiency and productivity, thereby reducing the operator's workload.

[0022] A configuration of a factor analysis system including a factor analysis apparatus 100 will be described. FIG. 2 is a diagram illustrating an example configuration of a factor analysis system 10 including a factor analysis apparatus 100A according to a first embodiment of the present disclosure. FIG. 3 is a flowchart illustrating an example of processing by the factor analysis apparatus 100, 100A of the present disclosure. FIG. 4 is a diagram illustrating a first example of causal relationship data used in the factor analysis apparatus 100, 100A of the present disclosure. FIG. 5 is a diagram illustrating a second example of causal relationship data used in the factor analysis apparatus 100, 100A of the present disclosure. FIG. 6 is a diagram illustrating a third example of causal relationship data used in the factor analysis apparatus 100, 100A of the present disclosure.

[0023] The factor analysis system 10 includes a factor analysis device 100A, a sensor 200, and a display device 300. The factor analysis device 100A uses a causal graph (causal relationship data) relating to the causal or dependency relationships between each target value (objective variable) and monitoring data (explanatory variables) to quantify the magnitude of the impact of the monitoring data on the target value, thereby supporting the consideration and prioritization of measures to improve resource consumption efficiency and productivity.

[0024] The sensor 200 is provided in the production facility that is the monitored facility. The sensor 200 outputs sensor data to the factor analysis device 100. The sensor data is, for example, time-series data of sensor measurement values ​​for a predetermined period of time obtained at predetermined intervals by the sensor 200 provided in the target facility. The sensor data indicates, for example, at least one sensor measurement value of temperature, humidity, electric energy, gas volume, oil volume, water volume, or steam flow rate. Note that this is merely an example, and the sensor data may also include, for example, a command value or a control value such as a reference value for a predetermined period of time obtained at predetermined intervals by multiple sensors 200.

[0025] The display device 300 is, for example, a display device. The display device 300 displays the factor analysis results output by the factor analysis device 100. The display device 300 may be an input / output device configured to include an input device. In this case, the display device 300 is configured as an information processing terminal or the like.

[0026] An example configuration of the factor analysis device 100, 100A will be described. The factor analysis device 100, 100A includes a data acquisition unit 110, 110A, a variable setting unit 120, 120A, a model learning unit 130, 130A, a causal relationship data acquisition unit 140, 140A, a contribution degree calculation unit 150, 150A, a factor analysis unit 160, 160A, and a factor analysis result output unit 170, 170A.

[0027] The data acquisition units 110, 110A acquire sensor data and production management information as monitoring data. The data acquisition units 110, 110A acquire multiple types of monitoring data related to the monitoring target. The data acquisition units 110, 110A acquire multiple time-series sensor data collected by multiple sensors 200 installed in the target equipment and production management information related to production of the target equipment as monitoring data. The data acquisition units 110, 110A acquire sensor data from the sensors 200. The data acquisition units 110, 110A also acquire production management information related to production of the target equipment. The production management information is time-series data for a predetermined period of time obtained at predetermined intervals. The production management information is, for example, at least one of data related to production conditions (production speed, product replacement, workers, etc.), production quantity, and work time (value production time, defect response time, setup time, etc.). The sensor data and production management information are acquired as monitoring data. That is, the data acquisition units 110 and 110A acquire, as the monitoring data, multiple time-series sensor data collected by multiple sensors installed in the monitored equipment and production management information related to the production of the monitored equipment. Here, if the sampling periods differ depending on the data, resampling processing may be performed based on the longest sampling period among the monitoring data. For example, if the number of monitoring data is represented by n, the monitoring data is represented as X1, X2, X3, X4, ..., Xn. Furthermore, if data is collected at time 1, time 2, ..., time t, the monitoring data is represented as a two-dimensional data frame with the number of times t as rows and the number n as columns. The value at time 1 of the first monitoring data X1 is represented as X11, the value at time 1 of the second monitoring data X2 is represented as X21, and the value at time 2 of the first monitoring data X1 is represented as X12.

[0028] The variable setting units 120 and 120A acquire monitoring data and, based on the monitoring data, generate training monitoring data, factorial analysis monitoring data, a training dataset consisting of explanatory variables and objective variables, and a factorial analysis dataset. Based on the monitoring data, the variable setting units 120 and 120A generate training monitoring data, factorial analysis monitoring data, a training dataset consisting of explanatory variables and objective variables, and a factorial analysis dataset. The variable setting units 120 and 120A set one piece of data representing the consumption efficiency or productivity of some resource as the objective variable. Here, the objective variable may be the monitoring data selected by the user as is. The variable setting units 120 and 120A may, for example, set data representing product quality as the objective variable. The variable setting units 120 and 120A may use data obtained by preprocessing at least one piece of monitoring data selected by the user as the objective variable. For example, the variable setting units 120 and 120A may calculate a basic unit from the production quantity and resource consumption amount and set it as the objective variable. The dependent variable may be discretized using one or more thresholds (e.g., when the monitoring data is continuous). In this case, the threshold may be manually set by the user, or a statistical quantity such as the mean, median, or standard deviation of the dependent variable may be used. Furthermore, the variable setting unit 120, 120A selects two or more pieces of data from the monitoring data that are related to the dependent variable, i.e., considered to be potential factors that cause fluctuations in the dependent variable, and sets these as explanatory variables. However, the explanatory variables do not include monitoring data that is affected by the dependent variable. For example, in a causal graph described below, the variable setting unit 120, 120A does not include monitoring data that is on the result side (child, downstream) of the dependent variable as explanatory variables. The variable setting unit 120, 120A then generates a dataset that pairs the dependent variable and the explanatory variable. The variable setting unit 120, 120A extracts datasets corresponding to some or all of the times at which the datasets were generated and outputs them to the model learning unit 130, 130A as training datasets.Furthermore, when the variable setting units 120 and 120A extract data sets corresponding to some of the times as learning data sets, they output data sets corresponding to times other than the certain times as factor analysis data sets to the factor analysis result output units 170 and 170A. When the variable setting units 120 and 120A designate all of the data sets as learning data sets, they output all of the data sets as factor analysis data sets to the factor analysis result output units 170 and 170A. When the variable setting units 120 and 120A extract data sets corresponding to some of the times as learning data sets, they output monitoring data corresponding to the certain times as learning monitoring data and monitoring data corresponding to times other than the certain times as factor analysis monitoring data to the contribution calculation units 150 and 150A. When all of the data sets are designated as learning data sets, they output all of the monitoring data as factor analysis monitoring data and learning monitoring data to the contribution calculation units 150 and 150A.

[0029] The model training units 130 and 130A generate a trained model based on a training dataset, which receives explanatory variables as input and outputs a dependent variable. The model training units 130 and 130A generate a training model based on the training dataset. Specifically, when the explanatory variables of the training dataset are input, the model training units 130 and 130A optimize the parameters constituting the training model so that the values ​​output from the training model approach the dependent variable of the training dataset. The model training units 130 and 130A can use the following well-known training models: When the dependent variable is a continuous value, the model training units 130 and 130A can use, for example, a regression tree, a random forest regression, a neural network, a support vector regression, etc. When the dependent variable is a discrete value, the model training units 130 and 130A can use, for example, a decision tree, a random forest, a neural network, a support vector machine, etc. The training model may be configured by combining at least one or more methods.

[0030] The causal relationship data acquisition units 140, 140A acquire causal relationship data indicating causal relationships among multiple types of monitoring data. The causal relationship data is data indicating dependencies between data. The causal relationship data is data indicating causal relationships and dependencies between multiple types of monitoring data. The causal relationship data indicates causal relationships and dependencies between input data, which is data input to a learning model. The causal relationship data is data manually created by a user. Alternatively, the causal relationship data is data estimated from the monitoring data. Alternatively, the causal relationship data is data that is given in advance using both data manually created by a user and data estimated from the monitoring data. The causal relationship data is data that is expressed, for example, in a causal graph ( FIG. 4 , FIG. 5 , or FIG. 6 ) described below. In the following description, the causal relationship data will also be referred to as a causal graph.

[0031] The contribution calculation unit 150, 150A calculates the contribution, which is the degree of influence that an explanatory variable has on a target variable, based on causal relationship data indicating causal relationships in multiple types of monitoring data, the learning model, the learning monitoring data, and the factorial analysis monitoring data. The contribution calculation unit 150, 150A acquires a causal graph, a learning model, the learning monitoring data, and the factorial analysis monitoring data, and calculates a contribution that quantifies the influence that an explanatory variable has on a target variable for each data sample. Here, the causal graph may be stored in a causal graph storage unit in advance, for example, and output from the causal graph storage unit to the contribution calculation unit 150, 150A. If the contribution calculation unit 150, 150A is configured to directly receive the causal graph from the causal graph storage unit, the configuration may not include the causal relationship data acquisition unit. The contribution calculation units 150 and 150A calculate the contribution, which indicates the degree of influence that the explanatory variables have on the target variable for each data sample, based on the input causal graph, learning model, learning monitoring data, and factor analysis monitoring data.

[0032] The contribution calculation units 150 and 150A are an extension of conventional contribution calculation methods, taking into account causal graphs. The contribution calculation units 150 and 150A have the function of calculating the magnitude of the influence of explanatory variables on a target variable based on the inter-data dependency relationship shown in a causal graph, while also taking into account indirect influences. For example, methods for quantifying the magnitude of influence include SHAP values ​​and XAI (Explainable AI) such as LIME. The contribution calculation units 150 and 150A calculate the contribution of explanatory variables to a target variable using a trained learning model, factor analysis monitoring data, learning monitoring data, and a causal graph (causal relationship data). The contribution calculated by the contribution calculation units 150 and 150A is a contribution that takes into account not only direct contributions but also indirect contributions based on the causal graph (causal relationship data). For example, the contribution is a SHAP value that can be calculated using a learning model. The SHAP value of a given explanatory variable in a given data sample is a value equivalent to the difference between the output value of the learning model when the explanatory variable is included in the input of the learning model and the output value of the learning model when the explanatory variable is not included. Here, a data sample refers to data corresponding to a given time. The sum of the SHAP values ​​of all explanatory variables in a given data sample is equal to the difference between the output value of the learning model for the given data sample and the average output value of the learning model for all data samples. Therefore, the SHAP value is a value that quantifies the effect of the explanatory variable on the objective variable for each data sample, i.e., the extent to which the objective variable changes depending on the presence or absence of the explanatory variable.

[0033] In detail, for example, the SHAP value is expressed by the following formulas (1) and (2): The contribution calculation units 150 and 150A use the SHAP value to calculate the contribution based on the difference in the expected value of the learning model output value between the presence and absence of the explanatory variable i. The term A in formula (2) indicates the learning model f. In the calculation of v(S) shown in formula (2), the input to the learning model f is the explanatory variables selected by the user. Furthermore, the probability distribution P (term B in formula (2)) that takes into account the dependency is always defined based on all monitoring data including those other than the explanatory variables. Therefore, whether the explanatory variables are narrowed down in advance by the user's judgment or the explanatory variables are narrowed down based on the analysis results, factor analysis can be performed while taking into consideration the dependency between the monitoring data. In other words, it is possible to narrow down the factors making use of the user's knowledge. For example, when there are five explanatory variables, the relationship between the output value of the learning model and the SHAP value is expressed in the following formula (3). However, when "Φ" in formula (3) is ０ " is the average value of the output values ​​of the learning model, and "x" is all the explanatory variables.

[0034] When the explanatory variables included in the set S are given, the contribution calculation units 150 and 150A define the probability distribution P (term B in equation (2)) of the explanatory variables for calculating the value function v(S) based on the causal graph. The causal graph is assumed to be given. It may be created manually by a user, or it may be estimated from data using a known causal graph estimation method (PC algorithm, GES, LiNGAM, etc.), or both may be used. The causal graph is assumed to be given in the following format. Here, the causal graph may be stored in advance in a causal graph storage unit, for example.

[0035] The causal graph will now be described. FIG. 4 is a diagram showing a first example of causal relationship data (causal graph) used in the factor analysis apparatus 100, 100A of the present disclosure. FIG. 5 is a diagram showing a second example of causal relationship data (causal graph) used in the factor analysis apparatus 100, 100A of the present disclosure. FIG. 6 is a diagram showing a third example of causal relationship data (causal graph) used in the factor analysis apparatus 100, 100A of the present disclosure. The causal graph is assumed to be a directed acyclic graph (DAG) representing the relationship between monitoring data. The causal graph is specified in the form of, for example, a graph, a matrix, or a parent-child relationship. Graph: Nodes represent monitoring data Xi, and arrows represent the relationship between monitoring data Xi and Xj. The starting point of the arrow represents parent monitoring data Xi, and the end point represents child monitoring data Xj, representing a parent-to-child relationship. (Figure 4) Matrix: Rows represent parent monitoring data Xi, and columns represent child monitoring data Xj. If a parent-to-child relationship exists, the element in row i and column j of the matrix is ​​set to 1, and if no relationship exists, the element is set to 0. (Figure 5) Parent-child relationship: The set of parent monitoring data of child monitoring data Xj is designated as pa(Xj) to represent the parent-to-child relationship. (Figure 6) For example, for five monitoring data X1, ..., X5, causal graphs are specified as shown in Figures 4, 5, and 6. Here, Figures 4, 5, and 6 all represent the same causal graph.

[0036] For example, when a causal graph representing the relationship between monitoring data is given, the contribution calculation units 150 and 150A express the probability distribution P of the explanatory variables (term B in equation (2)) as shown in the following equations (4) and (5). The C term in equation (4) indicates the relationship between the monitored data. In the relationship between monitored data, the result (child) depends only on the cause (parent). (Once a parent is given, everything other than the child and parent is independent.) For example, suppose that the causal graph is given as follows for five monitored data X1, ..., X5. pa(X1) = Φ pa(X2) = Φ pa(X3) = {X1, X2} pa(X4) = {X2} pa(X5) = {X3, X4} For example, if the objective variable is X5 and the explanatory variables are X1, X2, and X4, that is, N = {1, 2, 4}, S∈{1, 4}, the probability distribution P (the B term in equation (2)) can be expressed as follows: P(X ２ │x１ , x ４ , x ３ , x ５ ) = P(X ２ ) ... (6) For example, if the objective variable is X5, the explanatory variables are X1, X2, and X4, that is, N = {1, 2, 4}, and S ∈ Φ, then P (the B term in equation (2)) can be expressed as follows: P(X １ , X ２ , X ４ │x ３ , x ５ ) = P(X １ ) P(X ２ ) P(X ４ │X ２ ) ... (7) For example, if the objective variable is X5, the explanatory variables are X1, X3, and X4, that is, N = {1, 3, 4}, and S ∈ Φ, then P (the B term in equation (2)) can be expressed as follows: P(X １ , X ３ , X ４ │x ２ , x ５ ) = P(X １ ) P(X ３ │X １ , x ２ ) P(X ４ │x ２ ) ... (8)

[0037] The contribution calculation units 150 and 150A can use, for example, a normal distribution, a Dirichlet distribution, a t-distribution, an empirical distribution, or the like as the probability distribution P (the B term in equation (2)). When a distribution having parameters is used as the probability distribution, each parameter can be determined using the training monitoring data. For example, when a normal distribution is used as the probability distribution, the expected value and the variance-covariance matrix can be calculated using the training monitoring data.

[0038] The contribution calculation unit 150, 150A calculates the contribution of the explanatory variable i of the n-th data sample of the factor analysis monitoring data as Φin, and outputs it to the factor analysis unit 160, 160A and the factor analysis result output unit 170, 170A.

[0039] The factor analysis unit 160, 160A acquires and aggregates the contribution degrees and calculates the factor degree. The factor analysis unit 160, 160A calculates the factor degree, which is the degree to which each explanatory variable is a factor, based on the contribution degrees. The factor degree calculated for each explanatory variable can also be considered a value indicating the degree to which the explanatory variable is likely to be a factor affecting the target variable. FIG. 7 is a diagram illustrating the contribution degrees output by the contribution degree calculation unit 150, 150A in the factor analysis device 100, 100A of the present disclosure. In detail, the contribution degree Φin calculated for each data sample of explanatory variable i is aggregated to calculate the factor degree Ci of explanatory variable i (FIG. 7). For example, the average absolute value of Φin can be used as the factor degree Ci of explanatory variable i. The factor analysis units 160 and 160A output the calculated factor degrees to the factor analysis result output units 170 and 170A.

[0040] The factor analysis result output unit 170, 170A outputs the factor analysis result based on the factor analysis dataset, contribution degree, and factor degree. The factor analysis result output unit 170, 170A acquires the factor analysis dataset, contribution degree, and factor degree, and outputs information related to the factor analysis result. FIG. 8 is a diagram illustrating the factor degree output by the factor analysis unit 160, 160A in the factor analysis device 100, 100A of the present disclosure. The factor analysis result output unit 170, 170A outputs information related to the factor degree using the factor degree. FIG. 9 is a diagram illustrating a first display example of the factor analysis result output by the factor analysis result output unit 170, 170A in the factor analysis device 100, 100A of the present disclosure. FIG. 9 shows an example of an output screen related to the factor degree when the number of explanatory variables is six. Information indicating the explanatory variable i and information indicating the factor degree Ci are displayed. (FIG. 9) In the initial state, the explanatory variables are displayed in the order in which they were input or in a pre-stored order. A sort button is displayed to sort the results in descending order based on the factoriality. Pressing the sort button sorts the results in descending order, and the sort button turns black. Pressing the sort button again sorts the results in the order in which the explanatory variables were entered or stored. In the initial state, the results may be sorted in descending order of factoriality. In this case, the sort button turns black.

[0041] The factor analysis result output unit 170, 170A outputs information indicating the relationship between the contribution, explanatory variable, and dependent variable using the contribution and the factor analysis dataset. FIG. 10 is a diagram showing a second display example of the factor analysis results output by the factor analysis result output unit 170, 170A in the factor analysis device 100, 100A of the present disclosure. FIG. 10 shows an example of an output screen showing the relationship between the explanatory variable, contribution, and dependent variable for explanatory variable 1. A scatter plot is displayed with the explanatory variable value on the horizontal axis and the contribution on the vertical axis. Alternatively, the axes for the explanatory variable value and the contribution may be reversed to display a scatter plot with the contribution on the horizontal axis and the explanatory variable value on the vertical axis. In the scatter plot, a dashed line indicates a contribution of 0. In the scatter plot, one point corresponds to one data sample. The density of the points indicates the magnitude relationship between the dependent variable values ​​in the factor analysis dataset. Pressing the Expand button displays a list of explanatory variables, and selecting an explanatory variable displays a scatter plot corresponding to the selected explanatory variable. When the response variable is discretized data, that is, categorical data, the shape of the points may be changed instead of the density of the points.

[0042] In addition to the above configuration, the factor analysis apparatus 100, 100A may be configured to include a control unit (not shown) and a storage unit (not shown). The control unit (not shown) controls the entire factor analysis apparatus 100, 100A. For example, the control unit (not shown) starts up the factor analysis apparatus 100, 100A in accordance with an external command, or shuts down or puts the factor analysis apparatus 100, 100A into a sleep state. The storage unit (not shown) stores various data used by the factor analysis apparatus 100, 100A. For example, the storage unit (not shown) stores output (output data) from each component of the factor analysis apparatus 100, 100A, and outputs data requested by each component to the requesting component. This is also true for other embodiments.

[0043] 3 is a factor analysis method by the factor analysis device 100, 100A, and includes a data acquisition step in which a data acquisition unit 110, 110A of the factor analysis device 100, 100A acquires multiple types of monitoring data related to a monitoring target, a variable setting step in which a variable setting unit 120, 120A of the factor analysis device 100, 100A generates, based on the monitoring data, training monitoring data, factor analysis monitoring data, a training data set consisting of explanatory variables and a response variable, and a factor analysis data set, a model learning step in which a model learning unit 130, 130A of the factor analysis device 100, 100A generates, based on the training data set, a learning model trained to receive explanatory variables as input and output a response variable, and a causal relationship data acquisition unit 140, 140A of the factor analysis device 100, 100A generates a causal relationship data in the multiple types of monitoring data. a contribution calculation step in which a contribution calculation unit (150, 150A) of the factor analysis device (100, 100A) calculates a contribution, which is the degree of influence that an explanatory variable has on a target variable, based on the causal relationship data indicating the causal relationship in multiple types of monitoring data, the learning model, the learning monitoring data, and the factor analysis monitoring data; a factor analysis step in which a factor analysis unit (160, 160A) of the factor analysis device (100, 100A) calculates a factor degree, which is the degree to which each explanatory variable is a factor, based on the contribution degree; and a factor analysis result output step in which a factor analysis result output unit (170, 170A) of the factor analysis device (100, 100A) outputs a factor analysis result based on the factor analysis dataset, the contribution degree, and the factor degree. In addition, if the contribution calculation unit 150, 150A is configured to directly receive the causal graph from the causal graph storage unit, the factor analysis method may include a step in which the contribution calculation unit 150, 150A receives the causal graph from the causal graph storage unit instead of the causal relationship data acquisition step.

[0044] 3, the factor analysis device 100, 100A first executes a monitoring data acquisition process. (Step ST110) In the monitoring data acquisition process, the data acquisition unit 110, 110A of the factor analysis device 100, 100A acquires multiple types of monitoring data related to the monitoring target. The data acquisition unit 110, 110A acquires sensor data and production management information from the sensor 200 as monitoring data and outputs the data to the variable setting unit 120, 120A.

[0045] Next, the factor analysis device 100, 100A executes a variable setting process. (Step ST120) In the variable setting process, the variable setting unit 120, 120A of the factor analysis device 100, 100A generates training monitoring data, factor analysis monitoring data, a training data set, and a factor analysis data set based on the monitoring data. The variable setting unit 120, 120A performs the variable setting process on the monitoring data output by the data acquisition unit 110, 110A in step ST110, and outputs the training data set to the model learning unit 130, 130A, the training monitoring data and factor analysis monitoring data to the contribution calculation unit 150, 150A, and the factor analysis data set to the factor analysis result output unit 170, 170A.

[0046] Next, the factor analysis device 100, 100A executes a model learning process. (Step ST130) In the model learning process, the model learning unit 130, 130A of the factor analysis device 100, 100A generates a learning model that is trained to input explanatory variables and output a dependent variable based on the learning data set. The model learning unit 130, 130A trains the learning data set output by the variable setting unit 120, 120A in step ST120 so that the explanatory variables are input and the dependent variable is output, and outputs the learning model to the contribution calculation unit 150, 150A.

[0047] Next, the factor analysis device 100, 100A executes a causal relationship data acquisition process (Step ST140). In the causal relationship data acquisition process, the causal relationship data acquisition unit 140, 140A of the factor analysis device 100, 100A acquires causal relationship data indicating causal relationships in multiple types of monitoring data.

[0048] Next, the factor analysis device 100, 100A executes a contribution calculation process. (Step ST150) In the contribution calculation process, the contribution calculation unit 150, 150A of the factor analysis device 100, 100A calculates a contribution, which is the degree of influence that the explanatory variables have on the target variable, based on causal relationship data indicating causal relationships in multiple types of monitoring data, the learning model, the learning monitoring data, and the factor analysis monitoring data. The contribution calculation unit 150, 150A acquires a causal graph and performs a contribution calculation process on the learning model output by the model learning unit 130, 130A in step ST130, the learning monitoring data output by the variable setting unit 120, 120A in step ST120, and the factor analysis monitoring data, and outputs the contribution to the factor analysis unit 160, 160A and the factor analysis result output unit 170, 170A.

[0049] Next, the factor analysis device 100, 100A executes factor analysis processing. (Step ST160) In the factor analysis processing, the factor analysis unit 160, 160A of the factor analysis device 100, 100A calculates a factor degree, which is the degree to which each explanatory variable is a factor, based on the contribution degree. The factor analysis unit 160, 160A performs factor analysis processing on the contribution degree output by the contribution degree calculation unit 150, 150A in step ST150, and outputs the factor degree to the factor analysis result output unit 170, 170A.

[0050] Next, the factor analysis device 100, 100A executes a factor analysis result output process. (Step ST170) In the factor analysis result output process, the factor analysis result output unit 170, 170A of the factor analysis device 100, 100A outputs a factor analysis result based on the factor analysis dataset, the contribution degree, and the factor degree. The factor analysis result output unit 170, 170A acquires the factor analysis dataset output by the variable setting unit 120, 120A in step ST120, the contribution degree output by the contribution degree calculation unit 150, 150A in step ST150, and the factor degree output by the factor analysis unit 160, 160A in step ST160, and presents information related to the factor analysis result.

[0051] Next, the factor analysis device 100, 100A ends the process shown in FIG.

[0052] The factor analysis device of the present disclosure is configured as follows: a data acquisition unit that acquires multiple types of monitoring data related to a monitoring target; a variable setting unit that generates, based on the monitoring data, training monitoring data, factor analysis monitoring data, a training dataset consisting of explanatory variables and a dependent variable, and a factor analysis dataset; a model learning unit that generates, based on the training dataset, a learning model that is trained to input explanatory variables and output dependent variables; a contribution calculation unit that calculates a contribution degree, which is the degree of influence that explanatory variables have on the dependent variable, based on causal relationship data indicating causal relationships in the multiple types of monitoring data, the learning model, the training monitoring data, and the factor analysis monitoring data; a factor analysis unit that calculates a factor degree, which is the degree to which each explanatory variable is a factor, based on the contribution degree; and a factor analysis result output unit that outputs a factor analysis result based on the factor analysis dataset, the contribution degree, and the factor degree. This makes it possible to provide a factor analysis device that can improve the accuracy of factor analysis.

[0053] The factor analysis method of the present disclosure is configured as follows: A factor analysis method using a factor analysis device, comprising: a data acquisition step in which a data acquisition unit of the factor analysis device acquires multiple types of monitoring data related to a monitoring target; a variable setting step in which a variable setting unit of the factor analysis device generates, based on the monitoring data, training monitoring data, factor analysis monitoring data, a training dataset consisting of explanatory variables and a dependent variable, and a factor analysis dataset; a model learning step in which a model learning unit of the factor analysis device generates, based on the training dataset, a training model that is trained to take explanatory variables as input and output dependent variables; a contribution calculation step in which a contribution calculation unit of the factor analysis device calculates a contribution that is the degree of influence that an explanatory variable has on a dependent variable, based on causal relationship data indicating causal relationships in multiple types of monitoring data, the training model, the training monitoring data, and the factor analysis monitoring data; and a factor analysis step in which a factor analysis unit of the factor analysis device calculates a factor degree that is the degree to which each explanatory variable is a factor, based on the contribution degree. and a factor analysis result output step in which a factor analysis result output unit of the factor analysis device outputs a factor analysis result based on the factor analysis dataset, the contribution degree, and the factor degree. This makes it possible to provide a factor analysis method that can improve the accuracy of factor analysis.

[0054] The factor analysis system disclosed herein is configured as follows: a sensor; a data acquisition unit that acquires multiple types of monitoring data related to a monitoring target measured by the sensor; a variable setting unit that generates, based on the monitoring data, training monitoring data, factor analysis monitoring data, a training dataset consisting of explanatory variables and a dependent variable, and a factor analysis dataset; a model learning unit that generates, based on the training dataset, a learning model that is trained to input explanatory variables and output dependent variables; a contribution calculation unit that calculates a contribution degree, which is the degree of influence that an explanatory variable has on a dependent variable, based on causal relationship data indicating a causal relationship in multiple types of monitoring data, the learning model, the training monitoring data, and the factor analysis monitoring data; a factor analysis unit that calculates a factor degree, which is the degree to which each explanatory variable is a factor, based on the contribution degree; a factor analysis result output unit that outputs a factor analysis result based on the factor analysis dataset, the contribution degree, and the factor degree; and a display device that displays the factor analysis result output by the factor analysis result output unit. As a result, the present disclosure has an effect of providing a factor analysis system that can improve the accuracy of factor analysis.

[0055] A factor analysis program according to the present disclosure is configured as follows: The factor analysis program causes a computer to operate as a factor analysis device including: a data acquisition unit that acquires multiple types of monitoring data related to a monitoring target; a variable setting unit that generates, based on the monitoring data, training monitoring data, factor analysis monitoring data, a training dataset including explanatory variables and a dependent variable, and a factor analysis dataset; a model learning unit that generates, based on the training dataset, a learning model trained to input explanatory variables and output dependent variables; a contribution calculation unit that calculates a contribution degree, which is the degree of influence that explanatory variables have on the dependent variable, based on causal relationship data indicating causal relationships in multiple types of monitoring data, the learning model, the training monitoring data, and the factor analysis monitoring data; a factor analysis unit that calculates a factor degree, which is the degree to which each explanatory variable is a factor, based on the contribution degree; and a factor analysis result output unit that outputs a factor analysis result based on the factor analysis dataset, the contribution degree, and the factor degree. This makes it possible to provide a factor analysis program that can improve the accuracy of factor analysis.

[0056] The factor analysis device of the present disclosure is further configured as follows. In the factor analysis device, the data acquisition unit further acquires, as the monitoring data, a plurality of time-series sensor data collected by a plurality of sensors provided in the monitored equipment and production management information related to the production of the monitored equipment. This provides an effect of providing a factor analysis device that can improve the accuracy of factor analysis related to production equipment. Furthermore, the present disclosure provides the same effect as the above by applying the above configuration to the factor analysis method, the factor analysis system, or the factor analysis program.

[0057] The factor analysis apparatus of the present disclosure is further configured as follows. The factor analysis apparatus is further characterized in that the contribution calculation unit calculates the contribution using XAI (Explainable AI). This allows the present disclosure to provide a factor analysis apparatus capable of calculating the contribution using XAI, thereby further improving the accuracy of factor analysis. The present disclosure also provides a factor analysis apparatus that makes it easy for a user to interpret the factor analysis process. Furthermore, the present disclosure provides the same effect as the above by applying the above configuration to the factor analysis method, the factor analysis system, or the factor analysis program.

[0058] Embodiment 2. Embodiment 2 describes a mode that enables the evaluation results regarding the validity of explanatory variables to be reflected. In the description of Embodiment 2, content that is the same as the content already described in Embodiment 1 may be omitted as appropriate.

[0059] An example configuration of a factor analysis system including a factor analysis apparatus 100B will be described. Fig. 11 is a diagram showing an example configuration of a factor analysis system 10 including a factor analysis apparatus 100B according to a second embodiment of the present disclosure. Note that Fig. 11 shows only the configuration characteristic of the second embodiment compared to the configuration shown in the first embodiment. In the second embodiment, a configuration shown in the first embodiment but not shown in Fig. 11 may be added.

[0060] The factor analysis device of the factor analysis system may be configured to accept an evaluation of the validity of the explanatory variables from a user on an output screen regarding the factor degrees output by the factor analysis result output unit. The factor analysis system 10 is configured to include a factor analysis device 100B, a sensor 200, and a display device 300.

[0061] The factor analysis device 100B includes a data acquisition unit 110B, a variable setting unit 120B, a model learning unit 130B, a causal relationship data acquisition unit 140B, a contribution calculation unit 150B, a factor analysis unit 160B, and a factor analysis result output unit 170B. Of the above configuration, Fig. 11 shows the variable setting unit 120B, the model learning unit 130B, the causal relationship data acquisition unit 140B, the contribution calculation unit 150B, the factor analysis unit 160B, and the factor analysis result output unit 170B. The following mainly describes the contents not described in the above-mentioned embodiment.

[0062] The factor analysis result output unit 170B outputs information about at least one explanatory variable determined to be invalid by the validity evaluation to the variable setting unit 120B as explanatory variable removal information. FIG. 12 is a diagram illustrating an example display of the factor analysis result output by the factor analysis result output unit 170B in the factor analysis device 100B according to the second embodiment of the present disclosure. The factor analysis result output unit 170B, for example, provides a check box in the removal determination column of the output screen regarding the factor intensity, and accepts user input. For example, the check box is initially unchecked. An input operation (pressing or pressing) on ​​the unchecked check box changes the check box to a checked state. Furthermore, an input operation (pressing or pressing) on ​​the checked check box changes the check box to an unchecked state. For example, the explanatory variable corresponding to the checked check box can be used as explanatory variable removal information. In the figure, explanatory variables 3 and 4 are used as explanatory variable removal information. When at least one check box is checked and the user clicks the button to re-run the factor analysis, the factor analysis result output unit 170B outputs the information as explanatory variable removal information to the variable setting unit 120B and performs processing related to updating based on the user's evaluation of validity.

[0063] The variable setting unit 120B uses the explanatory variable removal information output by the factor analysis result output unit 170B and information on the monitoring data set as explanatory variables from among the monitoring data to newly set, as updated explanatory variables, the monitoring data that is not included in the "monitoring data included in the explanatory variable removal information" from among the "monitoring data set as explanatory variables." Here, it is assumed that the information on the monitoring data set as explanatory variables is stored in the variable setting unit 120B.

[0064] The factor analysis device 100B is configured to subsequently replace the explanatory variables with the updated explanatory variables and perform processing similar to that of embodiment 1, obtain the updated learning model, the updated contribution rate, and the updated factor rate, and output information related to the updated factor analysis.

[0065] The processing in the above configuration will be described. The factor analysis result output unit 170B outputs information regarding at least one explanatory variable determined to be inappropriate by the validity evaluation to the variable setting unit 120B as explanatory variable removal information. The variable setting unit 120B uses the explanatory variable removal information output by the factor analysis result output unit 170B and information on the monitoring data set as explanatory variables in the monitoring data to newly set, as updated explanatory variables, the monitoring data that is not included in the "monitoring data included in the explanatory variable removal information" among the "monitoring data set as explanatory variables." Thereafter, in the processing described in embodiment 1, the explanatory variables are replaced with updated explanatory variables, and processing similar to embodiment 1 is performed to obtain an updated learning model, updated contribution rate, and updated factor rate, and information on the updated factor analysis is output. Note that the above processing related to updating based on the user's validity evaluation may be performed repeatedly.

[0066] In the second embodiment, a configuration has been described in which the user evaluates the factor analysis results based on the information on the factor analysis output on the display device, examines the factor candidates, removes the factor candidates that are determined to be unnecessary, and performs the factor analysis process again, thereby improving the accuracy of the factor analysis.

[0067] The factor analysis apparatus of the present disclosure is further configured as follows. Compared to the factor analysis apparatus of the other embodiments, the factor analysis result output unit receives an evaluation of the validity of the explanatory variables based on the factor analysis results, and the variable setting unit sets the explanatory variables without using the monitoring data that is the explanatory variable that has been evaluated to be removed based on the result of the evaluation of the validity of the explanatory variables. This provides an effect of providing a factor analysis apparatus that can further improve the accuracy of factor analysis. Furthermore, the present disclosure provides the same effect as the above by applying the above configuration to the factor analysis method, the factor analysis system, or the factor analysis program.

[0068] Embodiment 3. In embodiment 3, a form is described in which results that satisfy aggregation conditions indicating conditions related to the objective variable and explanatory variables can be output. In the description of embodiment 3, content that is the same as the content already described in embodiment 1 and embodiment 2 may be omitted as appropriate.

[0069] Fig. 13 is a diagram illustrating a configuration example of a factor analysis system 10 including a factor analysis apparatus 100C according to a third embodiment of the present disclosure. Note that Fig. 13 illustrates only the configuration characteristic of the third embodiment, in contrast to the configurations illustrated in the first and second embodiments. In the third embodiment, the configurations illustrated in the first and second embodiments but not illustrated in Fig. 13 may be added. Fig. 14 is a diagram illustrating the aggregation conditions used in the factor analysis apparatus 100C according to the third embodiment of the present disclosure.

[0070] The factor analysis device 100C may be configured to calculate, in the factor analysis unit 160C, a factor degree for a contribution degree that satisfies a condition based on aggregation condition information indicating conditions related to the objective variable and explanatory variables input by the user.

[0071] The factor analysis system 10 includes a factor analysis device 100C, a sensor 200, and a display device 300. The factor analysis device 100C includes a data acquisition unit 110C, a variable setting unit 120C, a model learning unit 130C, a causal relationship data acquisition unit 140C, a contribution calculation unit 150C, a factor analysis unit 160C, a factor analysis result output unit 170C, and a tallying condition acquisition unit 180C. Of the above configuration, FIG. 13 shows the variable setting unit 120C, the contribution calculation unit 150C, the factor analysis unit 160C, the factor analysis result output unit 170C, and the tallying condition acquisition unit 180C. The following mainly describes the contents not described in the above-described embodiment.

[0072] The tallying condition acquisition unit 180C acquires tallying conditions. The tallying condition information indicating the tallying conditions can be provided in a format such as that shown in Fig. 14. The format shown in Fig. 14 is a format in which a lower limit and an upper limit are set for each objective variable and explanatory variable.

[0073] The factor analysis unit 160C acquires the aggregation conditions and outputs the conditional contribution degree, the conditional factor degree, and the conditional factor analysis dataset corresponding to the monitoring data indicated in the aggregation conditions in the factor analysis dataset so as to satisfy the aggregation conditions.

[0074] The factor analysis result output unit 170C outputs the factor analysis result using the conditional contribution degree, the conditional factor degree, and the conditional factor analysis data set output by the factor analysis unit 160C.

[0075] An example of processing by the factor analysis apparatus 100C will be described below. First, the tallying condition acquisition unit 180C of the factor analysis apparatus 100C acquires the tallying conditions.

[0076] Next, the factor analysis unit 160C acquires a counting condition and outputs a conditional contribution degree, a conditional factor degree, and a conditional factor analysis dataset corresponding to the monitoring data indicated by the counting condition among the factor analysis datasets so as to satisfy the counting condition. The factor analysis unit 160C of the factor analysis device 100C uses the counting condition information acquired by the counting condition acquisition unit 180C to calculate the factor degree for the contribution degree that satisfies a predetermined condition in a manner similar to that of embodiment 1. For example, the contribution degree that satisfies a predetermined condition can be determined by extracting data sample numbers in the factor analysis dataset that are greater than or equal to the lower limit and less than or equal to the upper limit for each of the dependent variable and the explanatory variable, and using these data sample numbers for all of the dependent variable and the explanatory variable. Furthermore, the factor analysis dataset corresponding to the data sample numbers extracted for all of the dependent variable and the explanatory variable can be determined as the factor analysis dataset that satisfies the predetermined condition. If only an upper limit or a lower limit is set, data sample numbers that are less than or equal to the upper limit or greater than or equal to the lower limit are extracted. If neither an upper limit nor a lower limit is set, all data sample numbers are extracted. For example, as shown in Fig. 14, the contribution degree corresponding to the data sample number that satisfies all of the following conditions is extracted: the dependent variable is 2 or more and 4 or less, the explanatory variable 1 is 0 or more, the explanatory variable 3 is 10 or less, the explanatory variable 4 is 0.5 or more and 0.7 or less, and the explanatory variable 5 is 5 or more.

[0077] Next, the factor analysis unit 160C acquires a counting condition and outputs a conditional contribution degree, a conditional factor degree, and a conditional factor analysis dataset corresponding to the monitoring data indicated by the counting condition among the factor analysis datasets so as to satisfy the counting condition. Specifically, the factor analysis unit 160C outputs the factor degree calculated from the contribution degree that satisfies the predetermined condition as the conditional factor degree, the contribution degree that satisfies the predetermined condition as the conditional contribution degree, and the factor analysis dataset that satisfies the predetermined condition as the conditional factor analysis dataset to the factor analysis result output unit 170C. More specifically, the factor analysis unit 160C calculates the contribution degree that satisfies the predetermined condition for the contribution degree output by the contribution degree calculation unit 150C, the factor analysis dataset output by the variable setting unit 120C, and the counting condition information, and outputs the calculated conditional factor degree, conditional contribution degree, and conditional factor analysis dataset to the factor analysis result output unit 170C.

[0078] The factor analysis result output unit 170C outputs the factor analysis result using the conditional contribution degree, the conditional factor degree, and the conditional factor analysis dataset output by the factor analysis unit 160C. Specifically, the factor analysis result output unit 170C replaces the contribution degree with the conditional contribution degree, the factor degree with the conditional factor degree, and the factor analysis dataset with the conditional factor analysis dataset, respectively, and performs the same processing as in the first embodiment.

[0079] In the third embodiment, conditions are set for the objective variable and the explanatory variables, and only the contributions under certain conditions are tallied to calculate the factoriality, thereby further improving the accuracy of the factor analysis.

[0080] The factor analysis device of the present disclosure is further configured as follows. Compared to the factor analysis device of the other embodiments, the factor analysis unit acquires a counting condition and outputs a conditional contribution degree, a conditional factor degree, and a conditional factor analysis dataset corresponding to the monitoring data indicated by the counting condition among the factor analysis datasets so as to satisfy the counting condition, and the factor analysis result output unit outputs a factor analysis result using the conditional contribution degree, the conditional factor degree, and the conditional factor analysis dataset output by the factor analysis unit. This provides an effect of providing a factor analysis device that can further improve the accuracy of factor analysis. Furthermore, the present disclosure provides the same effect as the above by applying the above configuration to the factor analysis method, the factor analysis system, or the factor analysis program.

[0081] Embodiment 4. Embodiment 4 describes a form in which the fitness of a learning model to monitoring data can be evaluated. In the description of Embodiment 4, content similar to that already described in Embodiments 1, 2, and 3 may be omitted as appropriate.

[0082] FIG. 15 is a diagram illustrating a configuration example of a factor analysis system 10 including a factor analysis device 100D according to a fourth embodiment of the present disclosure. Note that FIG. 15 illustrates only the configuration characteristic of the fourth embodiment, in contrast to the configurations illustrated in the first, second, and third embodiments. In the fourth embodiment, the configurations illustrated in the first, second, and third embodiments but not illustrated in FIG. 15 may be added. FIG. 16 is a diagram illustrating a first display example of the factor analysis result output by the factor analysis result output unit 170D in the factor analysis device 100D according to the fourth embodiment of the present disclosure. FIG. 17 is a diagram illustrating a second display example of the factor analysis result output by the factor analysis result output unit 170D in the factor analysis device 100D according to the fourth embodiment of the present disclosure. The factor analysis system 10 includes the factor analysis device 100D, a sensor 200, and a display device 300.

[0083] The factor analysis device 100D may include a model fitness evaluation unit 190D that evaluates the fitness of the learning model to the monitoring data. The factor analysis device 100D includes a data acquisition unit 110D, a variable setting unit 120D, a model learning unit 130D, a causal relationship data acquisition unit 140D, a contribution calculation unit 150D, a factor analysis unit 160D, a factor analysis result output unit 170D, a counting condition acquisition unit 180D, and a model fitness evaluation unit 190D. Of the above configuration, FIG. 15 shows the variable setting unit 120D, the model learning unit 130D, the factor analysis result output unit 170D, and the model fitness evaluation unit 190D. The following mainly describes the content not described in the above-described embodiment.

[0084] The model fitness evaluation unit 190D calculates the fitness of the learning model using the factor analysis dataset and the learning model.

[0085] The factor analysis result output unit 170D outputs the factor analysis result by further using the goodness of fit calculated by the model goodness of fit evaluation unit 190D.

[0086] An example of processing by the factor analysis device 100D will be described. The model fitness evaluation unit 190D of the factor analysis device 100D uses the factor analysis dataset output by the variable setting unit 120D and the learning model output by the model learning unit 130D to generate a learning model output value that is the output when the explanatory variables of the factor analysis dataset are input to the learning model. The model fitness evaluation unit 190D calculates the fitness using the generated learning model output value and the objective variable of the factor analysis dataset. The model fitness evaluation unit 190D outputs the generated learning model output value and the calculated fitness to the factor analysis result output unit 170D.

[0087] The factor analysis result output unit 170D of the factor analysis device 100D outputs information regarding the degree of fit using the degree of fit output by the model fit evaluation unit 190D. The factor analysis result output unit 170D also outputs information regarding the learning model output using the learning model output value output by the model fit evaluation unit 190D and the factor analysis dataset output by the variable setting unit 120D. The model fit evaluation unit 190D inputs the explanatory variables of the factor analysis dataset into the learning model to generate a learning model output value. The learning model output value is compared with the objective variable of the factor analysis dataset to calculate the degree of fit. Here, the degree of fit is determined using at least one accuracy evaluation index set by the user. When the objective variable is continuous, for example, the mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), coefficient of determination, etc. can be used. When the objective variable is binary categorical data, for example, accuracy, precision, recall, F1 value, etc. can be used. When the objective variable is ternary or more categorical data, for example, accuracy, average F1 value, etc. can be used. However, when the training dataset and the factorial analysis dataset are equal, the goodness of fit may be calculated by cross-validation. In this case, the training model used is not the training model output from the model training unit 130D, but a new training model trained for cross-validation. Here, cross-validation can be performed using methods such as leave-one-out cross-validation, k-fold cross-validation, and stratified k-fold cross-validation. When two or more goodness of fit values ​​are calculated for one accuracy evaluation index through cross-validation, for example, the average, maximum value, minimum value, median, etc., is used to calculate one representative value for one accuracy evaluation index. The model fitness evaluation unit 190D outputs the generated learning model output value and the fitness calculated using at least one or more accuracy evaluation indexes to the factor analysis result output unit 170D.

[0088] The factor analysis result output unit 170D outputs information related to the degree of fit using the degree of fit output by the model fit evaluation unit 190D. For example, assume that the types of accuracy evaluation indexes and the degrees of fit corresponding to the indexes are displayed in a table format as shown in Figure 16. Figure 16 shows a case where the objective variable is binary categorical data, and the F1 value, accuracy rate (Accuracy), and degree of fit (Precision) are selected by the user as the accuracy evaluation indexes.

[0089] The factor analysis result output unit 170D outputs information about the learning model output using the learning model output value output by the model fitness evaluation unit 190D and the factor analysis dataset output by the variable setting unit 120D. For example, as shown in FIG. 17 , a scatter plot is displayed with the data sample number of the factor analysis dataset on the horizontal axis and the dependent variable and model output value on the vertical axis. Instead of a scatter plot, a line graph may also be used. In FIG. 17 , black circles represent dependent variable values, and white circles represent learning model output values.

[0090] The fourth embodiment shows a configuration that can evaluate the fitness of a learning model. This improves the reliability of factor analysis by determining whether the learning model is good or bad. For example, if the fitness is poor, the settings of the explanatory variables can be revised.

[0091] The factor analysis device of the present disclosure is further configured as follows. Compared to the factor analysis device of the other embodiments, the factor analysis device further includes a model fitness evaluation unit that calculates the fitness of the learning model using the factor analysis dataset and the learning model, and the factor analysis result output unit outputs the factor analysis result by further using the fitness calculated by the model fitness evaluation unit. This provides an effect of providing a factor analysis device that can further improve the reliability of factor analysis. Furthermore, the present disclosure provides the same effect as the above by applying the above configuration to the factor analysis method, the factor analysis system, or the factor analysis program.

[0092] Here, a hardware configuration for realizing the functions of the configuration of the present disclosure described above will be described. FIG. 18 is a diagram illustrating a first example of a hardware configuration for realizing the functions of the configuration of the present disclosure. The hardware configuration illustrated in FIG. 18 realizes the functions of the factor analysis devices 100, 100A, 100B, 100C, and 100D. FIG. 19 is a diagram illustrating a second example of a hardware configuration for realizing the functions of the configuration of the present disclosure. The hardware configuration illustrated in FIG. 19 executes software that realizes the functions of the factor analysis devices 100, 100A, 100B, 100C, and 100D. The auxiliary storage device 1010 is a storage device having a storage area for reading and writing data. A storage unit (not illustrated) is realized by the auxiliary storage device 1010 or another memory (not illustrated). The information input device 1030 is a device for inputting data to the factor analysis devices 100, 100A, 100B, 100C, and 100D, and is, for example, a touch panel, a mouse, and a keyboard. Data input using the information input device 1030 is input to the factor analysis device via the information input IF 1020. The display IF 1040 is an IF that relays data output from the factor analysis devices 100, 100A, 100B, 100C, and 100D to the display 1050. The display 1050 displays the data. For example, the factor analysis result output units 170, 170A, 170B, 170C, and 170D output the results to the display 1050 via the display IF 1040. The display 1050 displays the estimation results input via the display IF 1040 on a screen. When the processing circuitry is dedicated hardware, the processing circuitry 1060 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.The processing circuit uses the above-described hardware to realize the functions of data acquisition units 110, 110A, 110B, 110C, 110D, variable setting units 120, 120A, 120B, 120C, 120D, model learning units 130, 130A, 130B, 130C, 130D, causal relationship data acquisition units 140, 140A, 140B, 140C, 140D, contribution calculation units 150, 150A, 150B, 150C, 150D, factor analysis units 160, 160A, 160B, 160C, 160D, factor analysis result output units 170, 170A, 170B, 170C, 170D, aggregation condition acquisition units 180C, 180D, model fitness evaluation unit 190D, and a control unit (not shown). When the processing circuit is a processor 1070, the functions of the factor analysis devices 100, 100A, 100B, 100C, and 100D are realized by software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 1080. The processor 1070 realizes the functions of the factor analysis devices 100, 100A, 100B, 100C, and 100D by reading and executing the program stored in the memory 1080. These programs cause a computer to execute each procedure or method of the functions of the factor analysis devices 100, 100A, 100B, 100C, and 100D. The processor 1070 executes software that realizes the functions of the data acquisition units 110, 110A, 110B, 110C, and 110D, the variable setting units 120, 120A, 120B, 120C, and 120D, the model learning units 130, 130A, 130B, 130C, and 130D, the causal relationship data acquisition units 140, 140A, 140B, 140C, and 140D, the contribution calculation units 150, 150A, 150B, 150C, and 150D, the factor analysis units 160, 160A, 160B, 160C, and 160D, the factor analysis result output units 170, 170A, 170B, 170C, and 170D, the aggregation condition acquisition units 180C and 180D, the model fitness evaluation unit 190D, and a control unit (not shown). The memory 1080 may be a computer-readable storage medium that stores a program for causing the factor analysis devices 100, 100A, 100B, 100C, and 100D to function.The memory 1080 may be, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), or an EEPROM (Electrically-EPROM), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a DVD.

[0093] It should be noted that some of the functions of the factor analysis devices 100, 100A, 100B, 100C, and 100D may be realized by dedicated hardware and some by software or firmware. The functions of the data acquisition units 110, 110A, 110B, 110C, 110D, the variable setting units 120, 120A, 120B, 120C, 120D, the model learning units 130, 130A, 130B, 130C, 130D, the causal relationship data acquisition units 140, 140A, 140B, 140C, 140D, the contribution calculation units 150, 150A, 150B, 150C, 150D, the factor analysis units 160, 160A, 160B, 160C, 160D, the factor analysis result output units 170, 170A, 170B, 170C, 170D, the aggregation condition acquisition units 180C, 180D, the model fitness evaluation unit 190D, and the control unit (not shown) may be realized by separate processing circuits, or may be realized together by a processing circuit.

[0094] or data acquisition units 110, 110A, 110B, 110C, 110D, variable setting units 120, 120A, 120B, 120C, 120D, model learning units 130, 130A, 130B, 130C, 130D, causal relationship data acquisition units 140, 140A, 140B, 140C, 140D, contribution degree calculation units 150, 150A, 150B, 150C, 150D, factor analysis units 160, 160A , 160B, 160C, 160D, factor analysis result output units 170, 170A, 170B, 170C, 170D, aggregation condition acquisition units 180C, 180D, model fitness evaluation unit 190D, and some of the functions of a control unit not shown may be realized by processor 10001 and memory 10002, and the remaining functions may be realized by processing circuit 20001.

[0095] Thus, the processing circuitry may implement each of the above functions through hardware, software, firmware, or a combination thereof.

[0096] It should be noted that the present disclosure allows for free combination of the embodiments, modification of any of the components of the embodiments, or omission of any of the components of the embodiments.

[0097] The present disclosure can improve the accuracy of factor analysis by calculating the contribution of each factor while taking into consideration the dependency relationships or causal relationships between multiple factors, and is therefore suitable for use in, for example, factor analysis devices, factor analysis methods, factor analysis systems, and factor analysis programs used in factor analysis of production equipment.

[0098] 10 Factor analysis system, 100, 100A, 100B, 100C, 100D Factor analysis device, 110, 110A, (110B, 110C, 110D) Data acquisition unit, 120, 120A, 120B, 120C, 120D Variable setting unit, 130, 130A, 130B, 130C, 130D Model learning unit, 140, 140A, 140B, (140C, 140D) Causal relationship data acquisition unit, 150, 150A, 150B, 150C, (150D) Contribution calculation unit, 160, 160A, 160B, 160C, (160D) Factor analysis unit, 170, 170A, 170B, 170C, 170D Factor analysis result output unit, 180C, (180D) aggregation condition acquisition unit, 190D model fitness evaluation unit, 200 sensor, 300 display device, 1010 auxiliary storage device, 1020 information input interface (information input IF), 1030 information input device, 1040 display interface (display IF), 1050 display, 1060 processing circuit, 1070 processor, 1080 memory.

Claims

a variable setting unit that generates, based on the monitoring data, monitoring data for learning, monitoring data for factor analysis, a learning dataset consisting of explanatory variables and a dependent variable, and a dataset for factor analysis; a model learning unit that generates, based on the learning dataset, a learning model that is trained to input explanatory variables and output dependent variables; a contribution calculation unit that calculates a contribution degree, which is the degree of influence that explanatory variables have on the dependent variable, based on causal relationship data indicating causal relationships in the multiple types of monitoring data, the learning model, the monitoring data for learning, and the monitoring data for factor analysis; a factor analysis unit that calculates a factor degree, which is the degree to which each explanatory variable is a factor, based on the contribution degree; and a factor analysis result output unit that outputs a factor analysis result based on the dataset for factor analysis, the contribution degree, and the factor degree.

2. The factor analysis device according to claim 1, characterized in that the data acquisition unit acquires, as the monitoring data, multiple time-series sensor data collected by multiple sensors installed in the monitored equipment, and production management information related to the production of the monitored equipment.

3. The factor analysis device according to claim 1 or 2, wherein the contribution degree calculation unit calculates the contribution degree using XAI (Explainable AI).

4. The factor analysis device according to any one of claims 1 to 3, wherein the factor analysis result output unit receives an evaluation of the validity of the explanatory variables based on the factor analysis results, and the variable setting unit sets the explanatory variables without using the monitoring data that is the explanatory variable that has been evaluated to be removed based on the result of the evaluation of the validity of the explanatory variables.

5. The factor analysis device according to any one of claims 1 to 4, wherein the factor analysis unit obtains a calculation condition and outputs a conditional contribution degree, a conditional factor degree, and a conditional factor analysis dataset corresponding to the monitoring data indicated in the calculation condition among the factor analysis datasets so as to satisfy the calculation condition, and the factor analysis result output unit outputs a factor analysis result using the conditional contribution degree, the conditional factor degree, and the conditional factor analysis dataset output by the factor analysis unit.

6. A factor analysis device as claimed in any one of claims 1 to 5, further comprising a model fitness evaluation unit that calculates a fitness of the learning model using the factor analysis dataset and the learning model, and the factor analysis result output unit outputs a factor analysis result by further using the fitness calculated by the model fitness evaluation unit.

7. A factor analysis method using a factor analysis device, comprising: a data acquisition step in which a data acquisition unit of the factor analysis device acquires multiple types of monitoring data related to a monitoring target; a variable setting step in which a variable setting unit of the factor analysis device generates, based on the monitoring data, monitoring data for learning, monitoring data for factor analysis, a learning dataset consisting of explanatory variables and target variables, and a dataset for factor analysis; a model learning step in which a model learning unit of the factor analysis device generates, based on the learning dataset, a learning model that is trained to input explanatory variables and output target variables; a contribution calculation step in which a contribution calculation unit of the factor analysis device calculates a contribution degree, which is the degree of influence that an explanatory variable has on a target variable, based on causal relationship data indicating causal relationships in multiple types of monitoring data, the learning model, the monitoring data for learning, and the monitoring data for factor analysis; and a factor analysis step in which a factor analysis unit of the factor analysis device calculates a factor degree, which is the degree to which each explanatory variable is a factor, based on the contribution degree. a factor analysis result output step in which a factor analysis result output unit of the factor analysis device outputs a factor analysis result based on the factor analysis data set, the contribution degree, and the factor degree.

Images