Information processing apparatus, information processing system, information processing method, and program

Abstract

To generate an appropriate structural causal model by detecting a difference in causal relation.SOLUTION: The information processing apparatus includes a processing unit. The processing unit generates a plurality of external noise estimation values corresponding to a plurality of variables for each of one or more pieces of result data based on the one or more pieces of result data including a plurality of result values each corresponding to the plurality of variables and a pre-update model that is a structural causal model representing a causal relationship of the plurality of variables. The processing unit determines whether or not a causal relationship between a plurality of variables represented by the one or more pieces of actual data is different from a causal relationship between the plurality of variables represented by the pre-update model based on independence between any two or more variables in the plurality of external noise estimation values for each of the one or more pieces of actual data. When it is determined that the causal relationship is different, the processing unit generates a post-update model, which is a structural causal model, based on the one or more pieces of actual data.SELECTED DRAWING: Figure 5

Description

[Technical Field]

[0001] The present invention relates to an information processing device, an information processing system, an information processing method, and a program. [Background technology]

[0002] Techniques have been proposed that use statistics and machine learning to unravel the complex structures hidden in data, such as causal discovery and causal inference, to analyze causal relationships from data and improve prediction and decision-making methods.

[0003] For example, in manufacturing systems, causal discovery and inference are used to identify factors that affect product quality and predict the impact of process changes on product quality based on raw material data, process data, quality inspection data, and maintenance data, etc. In information systems, causal relationships between system components are identified to perform fault analysis and performance improvement.

[0004] However, the causal structure of data in manufacturing systems and information systems is not constant over time and may change. When analyzing causal relationships in data using causal discovery and causal inference, it is desirable to update the model used for analysis, etc., as the causal structure changes. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] JP 2009-54843 A [Patent Document 2] Patent No. 5108116 [Patent Document 3] Patent No. 6937345 [Patent Document 4] Patent No. 7173395 [Patent Document 5] Patent No. 7251687 [License 6] Patent No. 5825599 [License 7] Special Announcement 2023-173391 [License 8] International Publication No. 2022 / 259446 [Non-licensed literature]

[0006] [Non-licensed Document 1] Peters, J., Janzing, D., & Scholkopf, B., “Elements of causal inference: foundations and learning algorithms” published in 2017, The MIT Press, pp.81-155 [Non-licensed Document 2] Shimizu, S., Hoyer, PO, Hyvarinen, A., Kerminen, A., & Jordan, M., “A linear non-Gaussian acyclic model for causal discovery”, 2006, Journal of Machine Learning Research, 7(10) [Non-licensed Document 3] Shimizu, S., Inazumi, T., Sogawa, Y., Hyvarinen, A., Kawahara, Y., Washio, T., & Hoyer, PO, "DirectLiNGAM: A direct method for learning a linear non-Gaussian structural equation model", 2011, Journal of Machine Learning Research-JMLR, 12(Apr), pp.1225-1248 [Non-licensed Document 4] Hyvarinen, A., & Smith, SM, “Pairwise likelihood ratios for estimation of non-Gaussian structural equation models”, 2013, The Journal of Machine Learning Research, 14(1), pp.111-152 [Non-patent document 5] Hoyer, P., Janzing, D., Mooij, JM, Peters, J., & Scholkopf, B., “Nonlinear causal discovery with additive noise models”, 2008, Advances in neural information processing systems, 21 [Non-patent document 6] Peters, J., Mooij, JM, Janzing, D., & Scholkopf, B., “Causal Discovery with Continuous Additive Noise Models”, published 2014, Journal of Machine Learning Research, 15, pp.2009-2053 [Non-Patent Document 7] Peters, J., & Buhlmann, P., “Identifiability of Gaussian structural equation models with equal error variances”, 2014, Biometrika, 101(1), pp.219-228 [Non-patent document 8] Vukovic, M., & Thalmann, S., “Causal discovery in manufacturing: A structured literature review”, 2022, Journal of Manufacturing and Materials Processing, 6(1), 10

Non-licensed literature 9

Non-licensed literature 10

Non-licensed Document 11

Non-licensed Document 12

Non-licensed Document 13

[0007] The problem to be solved by the present invention is to detect differences in causal relationships and generate an appropriate structural causal model. [Means for solving the problem]

[0008] An information processing device according to an embodiment includes a processing unit. The processing unit generates, for each of the one or more pieces of performance data, a plurality of external noise estimates corresponding to the plurality of variables, based on one or more pieces of performance data including a plurality of performance values ​​corresponding to the plurality of variables and a pre-update model that is a structural causal model representing the causal relationships among the plurality of variables. Each of the plurality of external noise estimates represents an estimate of an influence of external noise different from the influence of the plurality of variables on a corresponding variable among the plurality of variables. The processing unit determines, based on the independence between any two or more variables in the plurality of external noise estimates for each of the one or more pieces of performance data, whether the causal relationships among the plurality of variables represented by the one or more pieces of performance data differ from the causal relationships among the plurality of variables represented by the pre-update model. If the processing unit determines that the causal relationships differ, it generates an updated model that is the structural causal model based on the one or more pieces of performance data. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a configuration diagram of an information processing system according to an embodiment. [Figure 2] Diagram showing the structural causal model. [Figure 3] Diagram showing a causal graph. [Figure 4] FIG. 10 is a diagram showing an example of an adjacency matrix. [Figure 5] FIG. 1 is a diagram illustrating the configuration of a model updating device. [Figure 6] FIG. [Figure 7] FIG. 10 is a diagram showing an extrinsic noise matrix. [Figure 8] 10 is a flowchart showing the flow of processing by the model updating device. [Figure 9] FIG. 10 is a diagram showing an example of information output from the model updating device. [Figure 10] FIG. 10 is a diagram showing an example of an adjacency matrix of an updated model. [Figure 11] FIG. 10 is a diagram showing the difference matrix between the adjacency matrices before and after updating. [Figure 12] FIG. 10 is a bar graph showing the difference between the adjacency matrices before and after updating. [Figure 13] 10A and 10B are diagrams showing examples of displaying causal graphs before and after updating. [Figure 14] FIG. 10 is a diagram showing an example of a causal graph of an updated model. [Figure 15] FIG. 10 is a diagram showing a modified example of the information processing system according to the embodiment. [Figure 16] FIG. [Figure 17] FIG. 2 is a hardware configuration diagram of a model updating device. DETAILED DESCRIPTION OF THE INVENTION

[0010] 1 is a diagram showing the configuration of an information processing system 10 according to the embodiment. The information processing system 10 according to the embodiment includes a target system 20, a model storage device 50, and a model update device 60.

[0011] The target system 20 is, for example, a manufacturing system that manufactures products. The target system 20 may also be a data processing system that executes computer processes, or an information processing system that provides information processing services using information processing. The target system 20 is not limited to such systems, and may be any system that handles data.

[0012] The model updating device 60 is an information processing device that executes information processing. The model updating device 60 acquires one or more pieces of performance data including a plurality of performance values ​​each corresponding to a plurality of variables from the target system 20. The model updating device 60 updates the structural causal model (SCM) stored in the model storage device 50.

[0013] Each of the multiple variables represents a value sampled in the target system 20. For example, if the target system 20 is a manufacturing system, each of the multiple variables represents raw material data such as the amount and quality of raw materials, process data such as the operating time of equipment during manufacturing and sensor data detected by a sensor on the environment of the manufacturing equipment, quality inspection data indicating the quality of manufactured products, and maintenance data detected during maintenance.

[0014] Each of the one or more pieces of performance data is a vector including a plurality of performance values. In this embodiment, each of the one or more pieces of performance data is a vector including d variables (X1, X2, ..., X d ) is a d-dimensional vector containing d actual values ​​that correspond one-to-one to the

[0015] Furthermore, each of the one or more pieces of performance data includes a plurality of performance values ​​sampled from the target system 20 under conditions such as different times. For example, the first performance data and the second performance data among the one or more pieces of performance data are values ​​sampled from the target system 20 at different times. However, the multiple performance values ​​included in one piece of performance data are values ​​sampled under the same conditions such as the same time.

[0016] In this embodiment, the model updating device 60 acquires n samples of performance data (n is an integer equal to or greater than 1). Each of the one or more performance data in this embodiment is assigned an index that identifies the conditions, such as the time at which it was sampled.

[0017] The model storage device 50 stores a structural causal model for the target system 20. The model storage device 50 stores in advance a structural causal model obtained by learning actual values ​​of a plurality of variables sampled from the target system 20, for example, when the target system 20 starts operating, is initialized, or is shipped from the factory.

[0018] A structural causal model is information that represents the causal relationships among multiple variables. That is, a structural causal model is information that, for each pair of two variables included in the multiple variables, represents whether one variable in the pair affects the other variable and the degree of influence.

[0019] In this embodiment, the structural causal model is a linear model in which the magnitude of influence is expressed by real numbers, and is expressed using an adjacency matrix B. For each combination of two variables included in the plurality of variables, the adjacency matrix B represents the magnitude of influence from one variable to the other. In this embodiment, the number of variables is d, and the adjacency matrix B is expressed as a square matrix with d rows and d columns.

[0020] Each element included in adjacency matrix B contains a real value representing the magnitude of influence from a variable identified by a row (one variable) to a variable identified by a column (the other variable). The magnitude of influence from one variable to another variable may be positive, negative, or 0. 0 indicates that one variable has no influence on the other variable. Note that adjacency matrix B contains the magnitude of influence of pairs of variables where one variable and the other variable are the same. The magnitude of influence of pairs where one variable and the other variable are the same is included in the diagonal elements of adjacency matrix B and is 0. Note that the rows and columns of adjacency matrix B may be reversed from the example of this embodiment.

[0021] After the target system 20 is put into operation, the model updating device 60 determines, based on one or more pieces of performance data, whether the causal relationships among the multiple variables sampled from the target system 20 differ from the causal relationships among the multiple variables represented by the structural causal model (pre-update model) stored in the model storage device 50. For example, the model updating device 60 determines whether the causal relationships among the multiple variables sampled from the target system 20 have changed in response to deterioration or a change in the situation of the target system 20. If the model updating device 60 determines that the causal relationships among the multiple variables sampled from the target system 20 differ from the causal relationships among the multiple variables represented by the pre-update model, it generates a new structural causal model (updated model) obtained by learning one or more pieces of performance data. Then, the model updating device 60 updates the structural causal model stored in the model storage device 50 to the generated updated model.

[0022] As a result, the model updating device 60 can generate a structural causal model that appropriately reflects the causal relationships among multiple variables sampled from the target system 20 at the current time, even if the causal relationships among multiple variables in the target system 20 have changed over time or due to changes in operating conditions, etc.

[0023] Furthermore, if it is determined that the causal relationships of the multiple variables sampled from the target system 20 differ from the causal relationships of the multiple variables represented by the structural causal model (pre-update model) stored in the model storage device 50, it is considered that the causal relationships of the multiple variables currently sampled from the target system 20 have changed from the original causal relationships. Therefore, when the model updating device 60 determines that the causal relationships of the multiple variables sampled from the target system 20 differ from the causal relationships of the multiple variables represented by the pre-update model, it further determines the content of such change in the causal relationships and outputs the determination content. In this way, the model updating device 60 can support the analysis, improvement, etc. of the target system 20.

[0024] Figure 2 is a diagram of the structural causal model.

[0025] In this embodiment, the structural causal model is expressed as in equation (1).

number

[0026] k, j are the d variables (X1,X2,…,X d ) is an index that identifies one of the variables in X. k is a function of d variables (X1,X2,…,X d ) with index k (X k ) value. B jk is a real value. B jk is a function of d variables (X1,X2,…,X d ), the variable with index j (X j ) and the variable with index k (X k ) and the variable with index j (X j ) to the variable with index k (X k ) and P a (k) is the variable with index k (X k ) represents the set of indices of parent variables that directly influence the

[0027] E k is the variable with index k (X k ) represents the magnitude of exogenous noise. Exogenous noise is noise caused by external factors other than the influence of multiple variables.

[0028] When expressed as a matrix, the structural causal model is expressed as in equation (2).

number

[0029] E is the sum of d external noises (E1, E2, ..., E d ) where T on the right side represents the transpose matrix.

number

[0030] X is a set of d variables (X1, X2, …, X d ) is a vector containing

[0031] B T is the transposed matrix of the adjacency matrix B. As shown in equation (4), the adjacency matrix B has d×d elements whose values ​​are real numbers.

number

[0032] In addition, B T is sometimes called an adjacency matrix, but in this embodiment, B is used as the adjacency matrix.

[0033] A structural causal model can also be expressed as multiple equations, as shown on the left side of Figure 2. A structural causal model is also called a structural equation model (SEM).

[0034] FIG. 3 is a diagram illustrating a causal graph representing a structural causal model.

[0035] A structural causal model is represented by a directed causal graph. A causal graph is a set of d variables (X1, ..., X d ) in a one-to-one correspondence.

[0036] B is an element of the adjacency matrix B jk is the variable with index j in the causal graph (X j ) from the node corresponding to the variable with index k (X k ) represents the value corresponding to the directed edge to the node corresponding to

[0037] The causal graph is B jk is nonzero, the variable with index j (X j ) from the node corresponding to the variable with index k (X k) and the causal graph contains a directed edge to the node corresponding to B jk If is zero, then the variable with index j (X j ) from the node corresponding to the variable with index k (X k ) does not contain a directed edge to the node corresponding to

[0038] Also, E k is the variable with index k (X k ) represents the exogenous noise affecting the corresponding node.

[0039] FIG. 4 is a diagram illustrating an example of the adjacency matrix B. As shown in FIG.

[0040] When d=5, the adjacency matrix B is expressed as shown in Figure 4. Each element included in the adjacency matrix B represents the magnitude of the influence from one variable specified by the row to another variable specified by the column.

[0041] Such an adjacency matrix B is estimated based on one or more pieces of performance data using, for example, a causal search algorithm. For example, the adjacency matrix B may be estimated using an algorithm such as multiple regression or Adaptive Lasso after defining a causal order using domain knowledge.

[0042] Furthermore, for example, the adjacency matrix B may be estimated using known causal structure and covariance structure analysis based on one or more pieces of performance data. For example, the adjacency matrix B may be estimated using information regarding the presence or absence of causality between variables and a causality search algorithm based on one or more pieces of performance data. For example, the adjacency matrix B may be estimated using a causality search algorithm under assumptions such as non-Gaussianity, non-linearity, or homoscedasticity. Non-Gaussianity is described in Non-Patent Documents 2, 3, and 4. Non-linearity is described in Non-Patent Documents 5 and 6. Homoscedasticity is described in Non-Patent Document 7.

[0043] Fig. 5 is a diagram showing the configuration of the model updating device 60. In the description of Fig. 5, Figs. 6 to 7 will be referred to. Fig. 6 is a diagram showing an actual data matrix X'. Fig. 7 is a diagram showing an external noise matrix E'.

[0044] The model updating device 60 includes a data acquisition unit 62, a data storage unit 64, an estimation unit 66, a determination unit 68, a learning unit 70, an updated model storage unit 72, a control unit 74, a result output unit 76, and an update unit 78.

[0045] The data acquiring unit 62 acquires one or more pieces of performance data from the target system 20. In this embodiment, the data acquiring unit 62 acquires performance data including d performance values ​​corresponding to d variables, for n samples.

[0046] The data storage unit 64 stores one or more pieces of acquired performance data. In this embodiment, the data storage unit 64 stores a performance data matrix X' including n rows corresponding to n samples and d columns corresponding to d variables, as shown in Fig. 6. The elements included in the i-th row and j-th column of the performance data matrix X' are x, which are performance values ​​corresponding to the variable with index j among the d variables included in the sample with index i among the n samples. ij In the performance data matrix X', i is an integer greater than or equal to 1 and less than or equal to n, and j is an integer greater than or equal to 1 and less than or equal to d.

[0047] The estimation unit 66 calculates a plurality of external noise estimation values ​​corresponding to a plurality of variables for each of the one or more pieces of performance data stored in the data storage unit 64, based on the pre-update model and the one or more pieces of performance data stored in the data storage unit 64. The pre-update model is a structural causal model stored in the model storage device 50.

[0048] The plurality of exogenous noise estimates for each of the one or more performance data correspond one-to-one to the plurality of variables, and each of the plurality of exogenous noise estimates represents an estimate of the influence of the exogenous noise on a corresponding variable among the plurality of variables.

[0049] In this embodiment, the estimation unit 66 generates an extrinsic noise matrix E'. The extrinsic noise matrix E' includes a plurality of extrinsic noise estimated values ​​for each of one or more pieces of performance data.

[0050] For example, the extrinsic noise matrix E' includes n rows corresponding to n samples and d columns corresponding to d variables, as shown in Fig. 7. The element in the i-th row and j-th column of the extrinsic noise matrix E' is an extrinsic noise estimate e that estimates the extrinsic noise given to the variable with index j among the d variables included in the sample with index i among the n samples. ij Includes: e ij is expressed by a real number.

[0051] The estimation unit 66 calculates the external noise matrix E' using equation (5).

number

[0052] Here, I represents a unit matrix. That is, the estimation unit 66 calculates the external noise matrix E' by multiplying the result data matrix X' by a matrix (IB) obtained by subtracting the adjacent matrix B from the unit matrix I.

[0053] The determination unit 68 determines whether or not the causal relationship between the multiple variables represented by the one or more pieces of actual data differs from the causal relationship between the multiple variables represented by the pre-update model, based on the independence between any two or more variables in the multiple extrinsic noise estimated values ​​for each of the one or more pieces of actual data. In this embodiment, the determination unit 68 determines whether or not the causal relationship between the multiple variables represented by the one or more pieces of actual data differs from the causal relationship between the multiple variables represented by the pre-update model, based on the extrinsic noise matrix E' generated by the estimation unit 66.

[0054] Here, consider a case where the structural causal model satisfies the assumptions of the Linear Non-Gaussian Acyclic Model (LiNGAM). That is, consider a case where the structural causal model is linear, the exogenous noise is independent and follows a non-Gaussian distribution, and the causal graph obtained from the adjacency matrix B satisfies acyclicity. In this case, if the adjacency matrix B of the pre-update model is sufficiently close to the true adjacency matrix B representing the causal relationships of the actual data matrix X', each column of the extrinsic noise matrix E' will also be independent and follow a non-Gaussian distribution. On the other hand, if the adjacency matrix B of the pre-update model differs from the true adjacency matrix B representing the correlations of the actual data matrix X', each column of the extrinsic noise matrix E' will no longer be independent. In other words, if the generation model of the actual data matrix X' currently acquired from the target system 20 differs from the pre-update model stored in the model storage device 50, the independence of each column of the extrinsic noise matrix E' will decrease. The determination unit 68 utilizes such a property to determine whether the causal relationships of the multiple variables sampled in the target system 20 differ from the causal relationships of the multiple variables represented by the pre-update model.

[0055] For example, the determination unit 68 selects two columns corresponding to the two variables in the extrinsic noise matrix E' for each combination of two variables included in the plurality of variables. Then, the determination unit 68 selects one of the selected two columns E' for each combination of two variables included in the plurality of variables. j and the other column E k The independence index value, which is an index representing the independence between

[0056] The determination unit 68 calculates, as the independence index value, for example, any one of a correlation coefficient, a rank correlation coefficient, a mutual information, a pairwise likelihood ratio, a distance correlation, a Hilbert-Schmidt Independence Criterion (HSIC), a Maximal Information Coefficient (MIC), and a Randomized Dependence Coefficient (RDC), or a value obtained by combining two or more of these, between two columns corresponding to two variables in the extrinsic noise matrix E'. Note that a method for calculating the pairwise likelihood ratio is described in Non-Patent Document 4. A method for calculating the distance correlation is described in Non-Patent Document 10. A method for calculating the HSIC is described in Non-Patent Document 11. A method for calculating the MIC is described in Non-Patent Document 12. A method for calculating the RDC is described in Non-Patent Document 13. Furthermore, the determination unit 68 may calculate mutual information using a kernel method as the independence index value. Mutual information using a kernel method is described in Non-Patent Document 3.

[0057] The determination unit 68 calculates such an independence index value for all combinations of two variables included in the plurality of variables. Then, based on the calculated independence index value, the determination unit 68 determines whether or not the causal relationship between the plurality of variables represented by one or more pieces of performance data differs from the causal relationship between the plurality of variables represented by the pre-update model.

[0058] For example, for each combination of two variables included in the plurality of variables, the calculated independence index value is compared with a threshold value to determine whether the combination of the two variables is independent of each other. The determination unit 68 uses, for example, a preset value as the threshold value. The determination unit 68 may also use, as the threshold value, a value determined by a permutation test or a bootstrap test, or a value corrected for multiple comparisons with a preset value.

[0059] For example, the determination unit 68 determines that the combination of two variables is independent when the calculated independence index value is greater than a threshold value, and determines that the combination of two variables is not independent, that is, one variable in the combination of two variables depends on the other variable, when the calculated independence index value is equal to or less than a threshold value.

[0060] Next, the judgment unit 68 judges whether the causal relationship of the multiple variables represented by one or more actual data items differs from the causal relationship of the multiple variables represented by the pre-update model, for example, based on the comparison result between the independence index value for each combination of two variables included in the multiple variables and a threshold value.

[0061] For example, when the determination unit 68 determines that all combinations of two variables included in the plurality of variables are independent, it determines that the causal relationships are not different. On the other hand, when the determination unit 68 determines that at least one combination of two variables included in the plurality of variables is not independent, it determines that the causal relationships are different.

[0062] Furthermore, for example, when the determination unit 68 determines that a predetermined first ratio of combinations of two variables included in the plurality of variables is independent, the determination unit 68 may determine that the causal relationships are not different. On the other hand, when the determination unit 68 determines that a predetermined second ratio of combinations of two variables included in the plurality of variables is not independent, the determination unit 68 may determine that the causal relationships are different. The predetermined second ratio is a ratio obtained by subtracting the first ratio from 100%.

[0063] Furthermore, for example, the determination unit 68 may determine that the causal relationships are not different when it determines that all of a predetermined number of pairs selected from a plurality of variables are independent. On the other hand, the determination unit 68 may determine that the causal relationships are different when it determines that at least one combination of the predetermined number of pairs selected from a plurality of variables is not independent.

[0064] Furthermore, for example, when the determination unit 68 determines that a combination of a predetermined first ratio among a predetermined number of sets selected from a plurality of variables is independent, the determination unit 68 may determine that the causal relationships are not different. On the other hand, when the determination unit 68 determines that a combination of a predetermined second ratio among a predetermined number of sets selected from a plurality of variables is not independent, the determination unit 68 may determine that the causal relationships are different.

[0065] Furthermore, for example, the determination unit 68 integrates the independence index values ​​calculated for each combination of two variables included in the plurality of variables into one statistical value, and determines whether the causal relationship between the plurality of variables represented by one or more pieces of performance data differs from the causal relationship between the plurality of variables represented by the pre-update model based on the comparison result between the statistical value and a threshold value. The statistical value is, for example, the sum of all the independence indexes or the product of all the independence indexes. The determination unit 68 uses, for example, a preset value as the threshold value. Furthermore, the determination unit 68 may use, as the threshold value, a value determined by a permutation test or a value corrected for multiple comparisons with a preset value.

[0066] For example, when a correlation coefficient is used as the independence index, the determining unit 68 uses Bartlett's on-sphere test, permutation test, or bootstrap test to determine whether the statistical value of the correlation matrix is ​​equal to the statistical value of the identity matrix, and to determine whether the causal relationship between multiple variables represented by one or more actual data sets differs from the causal relationship between multiple variables represented by the pre-update model. As the statistical value, for example, the determinant of the correlation matrix, the logarithm of the determinant, the norm, the largest eigenvalue, or the trace is used.

[0067] The learning unit 70 acquires the determination result of whether the causal relationships differ or not, made by the determination unit 68. If the determination unit 68 determines that the causal relationships differ, the learning unit 70 performs a learning process based on the performance data matrix X' stored in the data storage unit 64, and generates an updated model that is a structural causal model.

[0068] In this embodiment, the learning unit 70 includes a first-level learning unit 82, a second-level learning unit 84, and a third-level learning unit 86. The first-level learning unit 82, the second-level learning unit 84, and the third-level learning unit 86 perform learning processes using mutually different learning models based on one or more pieces of performance data stored in the data storage unit 64, to generate an updated model. The learning unit 70 may be configured to include any one or any two of the first-level learning unit 82, the second-level learning unit 84, and the third-level learning unit 86.

[0069] If it is determined that the causal relationships differ, the first-level learning unit 82 estimates new values ​​corresponding to the non-zero elements in the adjacency matrix B of the pre-update model stored in the model storage device 50 using linear regression, based on one or more pieces of actual data stored in the data storage unit 64. Then, the first-level learning unit 82 generates an updated model by updating the values ​​of the non-zero elements in the adjacency matrix B of the pre-update model to the estimated new values.

[0070] The updated model generated by the first level learning unit 82 is expressed as in equation (6).

number

[0071] The first level learning unit 82 calculates B in equation (6). jk The first-level learning unit 82 estimates B in equation (6) by linear regression such as the least squares method based on the performance data matrix X'. If the difference in values ​​is estimated to be small, the first-level learning unit 82 uses regularized linear regression, such as Transfer Lasso shown in Non-Patent Document 14, in which the difference between the adjacency matrix B of the pre-update model and the adjacency matrix B' of the updated model is limited by a regularization term. jk ' may be estimated.

[0072] The first level learning unit 82 extracts a sequence of pairs of two variables that are not determined to be independent, that is, pairs of two variables that are determined to have a dependency relationship, and calculates B in equation (6) only for the extracted two or more variables. jk In this case, the first level learning unit 82 sets the values ​​corresponding to the unextracted variables in the adjacency matrix B' of the updated model to be the same as the values ​​of the adjacency matrix B of the pre-update model.

[0073] The first-level learning unit 82 can generate a structural causal model in accordance with the causal relationships among multiple variables represented by one or more pieces of actual data when the causal order and causal structure among multiple variables represented by one or more pieces of actual data match the causal structure and causal order among multiple variables represented by the pre-update model but have different influences. In other words, the first-level learning unit 82 can generate a structural causal model in accordance with the causal relationships among multiple variables represented by one or more pieces of current actual data when the relationship between zero and non-zero values ​​of each element included in the adjacency matrix B is the same but the values ​​are different.

[0074] If it is determined that the causal relationships are different, the second-level learning unit 84 sets the causal order to be the same as the adjacency matrix B of the pre-update model, and generates an updated model using regularized linear regression based on one or more pieces of actual data.

[0075] The updated model generated by the second level learning unit 84 is expressed as in equation (7).

number

[0076] The second level learning unit 84 calculates B in equation (7). jk The upper triangular element of X′ is estimated based on the performance data matrix X′. For example, the second level learning unit 84 uses Lasso or Adaptive Lasso to estimate B jkFurthermore, for example, the second-level learning unit 84 fixes the causality order of the adjacency matrix B of the pre-update model, and uses either the Adaptive Lasso, the Transfer Lasso, or the Adaptive Transfer Lasso to estimate the components of the upper triangle of B in equation (7). jk Alternatively, an initial estimate of each of the components in the upper triangle of B' may be estimated. The Adaptive Lasso is described in Non-Patent Document 15. The Adaptive Transfer Lasso is described in Non-Patent Document 16. Then, the second-level learning unit 84 generates an updated model using the estimated adjacency matrix B'.

[0077] The second level learning unit 84 extracts a sequence of pairs of two variables that are not determined to be independent, that is, pairs of two variables that are determined to have a dependency relationship, and calculates B in equation (7) only for the extracted two or more variables. jk In this case, the second-level learning unit 84 sets the values ​​corresponding to the unextracted variables in the adjacency matrix B' of the updated model to be the same as the values ​​of the adjacency matrix B of the pre-update model.

[0078] When the causal order in the multiple variables represented by one or more pieces of performance data matches the causal order in the multiple variables represented by the pre-update model but the influence and causal structure are different, the second-level learning unit 84 can generate a structural causal model in accordance with the causal relationships in the multiple variables represented by one or more pieces of performance data. That is, when the causal order in the adjacency matrix B is the same but the values, including zero and non-zero relationships, are different, the second-level learning unit 84 can generate a structural causal model in accordance with the causal relationships in the multiple variables represented by one or more pieces of current performance data.

[0079] When it is determined that the causal relationships differ, the third level learning unit 86 estimates a structural causal model including the causal order based on one or more pieces of performance data.

[0080] For example, the third-level learning unit 86 generates the updated model based on a causal search method based on non-Gaussianity. Furthermore, for example, the third-level learning unit 86 may generate the updated model based on ICA-LiNGAM or DirectLiNGAM. ICA-LiNGAM is described in Non-Patent Document 2. DirectLiNGAM is described in Non-Patent Document 3.

[0081] When estimating the adjacency matrix B', the third level learning unit 86 extracts a pair of two variables that are not determined to be independent, that is, a pair of two variables that are determined to have a dependency relationship, and calculates B' only for the extracted two or more variables. jk In this case, the third level learning unit 86 sets the values ​​corresponding to the unextracted variables in the adjacency matrix B' of the updated model to be the same as the values ​​of the adjacency matrix B of the pre-update model.

[0082] Such a third-level learning unit 86 can generate a structural causal model that follows the causal relationships among multiple variables represented by one or more current performance data, even if the causal relationships among multiple variables represented by one or more performance data differ from the causal relationships among multiple variables represented by the pre-update model in terms of the magnitude of influence, causal order, and causal structure.

[0083] The updated model storage unit 72 stores the updated models generated by the first level learning unit 82, the second level learning unit 84, and the third level learning unit 86.

[0084] The control unit 74 controls which of the first level learning unit 82, the second level learning unit 84, and the third level learning unit 86 is to generate a structural causal model. For example, when it is determined that the causal relationships differ, the control unit 74 selects one of the first level learning unit 82, the second level learning unit 84, and the third level learning unit 86 to generate a structural causal model.

[0085] Alternatively, the control unit 74 may cause the first level learning unit 82, the second level learning unit 84, and the third level learning unit 86 to generate structural causal models in this order.

[0086] In this case, the control unit 74 causes the first level learning unit 82 to generate a structural causal model, and then causes the estimation unit 66 to generate an external noise matrix E' using the structural causal model generated by the first level learning unit 82, and causes the determination unit 68 to re-determine whether the causal relationships differ.

[0087] Furthermore, when it is determined that the causal relationships differ for the structural causal model generated by the first-level learning unit 82, the control unit 74 secondly causes the second-level learning unit 84 to generate a structural causal model. Next, the control unit 74 causes the estimation unit 66 to generate an extrinsic noise matrix E' using the structural causal model generated by the second-level learning unit 84, and causes the determination unit 68 to re-determine whether the causal relationships differ.

[0088] When the control unit 74 determines that the causal relationships are different for the structural causal model generated by the second-level learning unit 84, it thirdly causes the third-level learning unit 86 to generate a structural causal model, and terminates the processing. Note that if the LiNGAM assumption is satisfied, each column of the extrinsic noise matrix E' in the structural causal model generated by the third-level learning unit 86 is independent. Therefore, even if the determination unit 68 is made to re-determine whether the causal relationships are different, it is determined that the causal relationships are not different for the structural causal model generated by the third-level learning unit 86.

[0089] The control unit 74 may generate the structural causal models in the order of the first level learning unit 82 → the third level learning unit 86. The control unit 74 may also generate the structural causal models in the order of the second level learning unit 84 → the third level learning unit 86.

[0090] The result output unit 76 outputs the determination result of whether or not the causal relationships among the multiple variables represented by one or more pieces of performance data differ from the causal relationships among the multiple variables represented by the pre-update model, made by the determination unit 68. For example, the result output unit 76 displays the determination result on a display device or the like.

[0091] Furthermore, if it is determined that the causal relationship between the multiple variables sampled from the target system 20 differs from the causal relationship between the multiple variables represented by the pre-update model stored in the model storage device 50, it is considered that the causal relationship between the multiple variables currently sampled from the target system 20 has changed from the original causal relationship.

[0092] Therefore, when structural causal models are generated sequentially in the order of the first-level learning unit 82, the second-level learning unit 84, and the third-level learning unit 86, the result output unit 76 may further determine the content of the change in the causal relationship based on the characteristics of the learning models at each level and the re-determination results of the differences in the causal relationships for the structural causal models generated by the learning models at each level. Then, the result output unit 76 may display the determination result indicating the content of the determined change in the causal relationship on, for example, a display device.

[0093] For example, if it is determined that the causal relationships are not different from those of the pre-update model originally stored in the model storage device 50, it is determined that the causal relationships of the current plurality of variables sampled from the target system 20 have not changed from the original causal relationships, or that they cannot be said to have changed. Therefore, if it is determined that the causal relationships are not different from those of the pre-update model originally stored in the model storage device 50, the result output unit 76 may cause the display device to display at least one piece of information such as "no change in causal influence," "no change in causal structure," and "no change in causal order."

[0094] A change in causal influence indicates a change in the magnitude of the influence between any two pairs of variables among multiple variables. A change in causal structure indicates a change in the structure of causal relationships among multiple variables. A change in causal order indicates a change in the causal order of multiple variables.

[0095] Furthermore, for example, when it is re-determined that the causal relationships of the structural causal model generated by the first-level learning unit 82 are not different, it is determined that the causal relationships of the current plurality of variables sampled from the target system 20 are the same as the original causal relationships, with no change in the positions of zeros and non-zeros in the adjacency matrix B, and the causal structure and causal order are the same, but there is a change in the values ​​included in the adjacency matrix B, and the causal influence is different. Therefore, when it is re-determined that the causal relationships of the structural causal model generated by the first-level learning unit 82 are not different, the result output unit 76 may display at least one piece of information from "change in causal influence," "no change in causal structure," and "no change in causal order" on the display device.

[0096] Furthermore, for example, when it is re-determined that the causal relationships of the structural causal model generated by the second-level learning unit 84 are not different, it is determined that the causal order of the causal relationships of the multiple variables currently sampled from the target system 20 is the same as that of the original causal relationships, but that the causal influence and causal structure are different. Therefore, when it is re-determined that the causal relationships of the structural causal model generated by the second-level learning unit 84 are not different, the result output unit 76 may display at least one of the following information on the display device: "there is a change in the causal influence," "there is a change in the causal structure," and "there is no change in the causal order."

[0097] Furthermore, for example, when the structural causal model generated by the third-level learning unit 86 is re-determined to have different causal relationships, it is determined that the causal relationships of the current plurality of variables sampled from the target system 20 are all different from the original causal relationships in terms of causal influence, causal structure, and causal order. Therefore, when the structural causal model generated by the second-level learning unit 84 is re-determined to have different causal relationships, the result output unit 76 may display at least one of the following information on the display device: "there is a change in causal influence," "there is a change in causal structure," and "there is a change in causal order."

[0098] The result output unit 76 may also cause the display device to display similar information when structural causal models are generated sequentially, for example, from the first-level learning unit 82 to the third-level learning unit 86. The result output unit 76 may also cause the display device to display similar information when structural causal models are generated sequentially, for example, from the second-level learning unit 84 to the third-level learning unit 86.

[0099] After the learning unit 70 has completed learning, the update unit 78 reads the generated structural causal model from the updated model storage unit 72 and stores it in the model storage device 50.

[0100] 8 is a flowchart showing the processing flow of the model updating device 60 when structural causal models are generated sequentially in the order of the first-level learning unit 82, the second-level learning unit 84, and the third-level learning unit 86. The model updating device 60 may execute the processing according to the flow shown in FIG.

[0101] First, in S11, the model updating device 60 acquires one or more pieces of performance data.

[0102] Next, in S12, the model updating device 60 generates a plurality of first external noise estimated values, which are a plurality of external noise estimated values ​​for each of the one or more pieces of performance data, based on the one or more pieces of performance data and the pre-update model.

[0103] Next, in S13, the model updating device 60 executes a first determination process to determine whether or not the causal relationship between the multiple variables represented by the one or more pieces of performance data differs from the causal relationship between the multiple variables represented by the pre-update model, based on the independence between any two variables in the multiple first exogenous noise estimates for each of the one or more pieces of performance data. If the model updating device 60 determines that the causal relationship does not differ (No in S13) through the first determination process, it proceeds to S14, and if the model updating device 60 determines that the causal relationship differs (Yes in S13) through the first determination process, it proceeds to S15.

[0104] In S14, the model updating device 60 outputs information of the contents "no change in causal influence," "no change in causal structure," and "no change in causal order." The model updating device 60 may output any one or two pieces of information of "no change in causal influence," "no change in causal structure," and "no change in causal order." The model updating device 60 ends this flow when it finishes processing S14.

[0105] In S15, the model updating device 60 estimates new values ​​corresponding to non-zero elements in adjacency matrix B of the pre-update model using linear regression based on one or more pieces of performance data. Then, the model updating device 60 generates a first structural causal model, which is a structural causal model, by updating the values ​​of the non-zero elements in adjacency matrix B of the pre-update model to the estimated new values. For example, the model updating device 60 generates the first structural causal model using the first-level learning unit 82.

[0106] Next, in S16, the model updating device 60 generates a plurality of second exogenous noise estimates, which are a plurality of exogenous noise estimates for each of the one or more pieces of actual data, based on the one or more pieces of actual data and the first structural causal model.

[0107] Next, in S17, the model updating device 60 executes a second determination process to determine whether or not the causal relationship between the multiple variables represented by the one or more pieces of performance data differs from the causal relationship between the multiple variables represented by the first structural causal model, based on the independence between any two variables in the multiple second exogenous noise estimates for each of the one or more pieces of performance data. If the model updating device 60 determines that the causal relationship does not differ by the second determination process (No in S17), it proceeds to S18, and if the model updating device 60 determines that the causal relationship does differ by the second determination process (Yes in S17), it proceeds to S20.

[0108] In S18, the model updating device 60 outputs information of the contents "there is a change in causal influence," "there is no change in the causal structure," and "there is no change in the causal order." The model updating device 60 may output any one or two pieces of information of "there is a change in causal influence," "there is no change in the causal structure," and "there is no change in the causal order."

[0109] Following S18, in S19, the model updating device 60 uses the first structural causal model as the updated model to update the structural causal model stored in the model storage device 50. After completing the process of S19, the model updating device 60 ends this flow.

[0110] In S20, the model updating device 60 generates a second structural causal model using regularized linear regression based on one or more pieces of performance data, with the causal order set to be the same as that of the adjacency matrix B of the pre-update model. For example, the model updating device 60 generates a first structural causal model using the second-level learning unit 84.

[0111] Next, in S21, the model updating device 60 generates a plurality of third external noise estimates, which are a plurality of external noise estimates for each of the one or more pieces of actual data, based on the one or more pieces of actual data and the second structural causal model.

[0112] Next, in S22, the model updating device 60 executes a third determination process to determine whether or not the causal relationship between the multiple variables represented by the one or more pieces of performance data differs from the causal relationship between the multiple variables represented by the second structural causal model, based on the independence between any two variables in the multiple third exogenous noise estimates for each of the one or more pieces of performance data. If the model updating device 60 determines that the causal relationship does not differ by the third determination process (No in S22), it proceeds to S23, and if the model updating device 60 determines that the causal relationship differs by the second determination process (Yes in S22), it proceeds to S25.

[0113] In S23, the model updating device 60 outputs information of the contents "there is a change in causal influence," "there is a change in causal structure," and "there is no change in causal order." The model updating device 60 may output any one or two pieces of information of "there is a change in causal influence," "there is a change in causal structure," and "there is no change in causal order."

[0114] Following S23, in S24, the model updating device 60 uses the second structural causal model as the updated model to update the structural causal model stored in the model storage device 50. After completing the process of S24, the model updating device 60 ends this flow.

[0115] In S25, the model updating device 60 generates a third structural causal model, which is a structural causal model including a causal order, based on one or more pieces of performance data. For example, the model updating device 60 generates the third structural causal model using the third-level learning unit 85.

[0116] Subsequently, in S26, the model updating device 60 outputs information of the contents "there is a change in causal influence," "there is a change in causal structure," and "there is a change in causal order." The model updating device 60 may output any one or two pieces of information of "there is a change in causal influence," "there is a change in causal structure," and "there is a change in causal order."

[0117] Following S26, in S27, the model updating device 60 uses the third structural causal model as the updated model to update the structural causal model stored in the model storage device 50. After completing the process of S26, the model updating device 60 ends this flow.

[0118] Note that the model updating device 60 may proceed to S25 without executing the processes of S20 to S24, if it is determined by the second determination process that the causal relationships differ (Yes in S17). This enables the model updating device 60 to generate structural causal models in the order of the first-level learning unit 82 → the third-level learning unit 86.

[0119] Furthermore, the model updating device 60 may proceed to S20 if it is determined by the first determination process that the causal relationships differ (Yes in S13) without executing the processes from S15 to S19. This enables the model updating device 60 to generate structural causal models in the order of the second-level learning unit 84 → the third-level learning unit 86.

[0120] As described above, the model updating device 60 generates a plurality of new structural causal models, each of which is a structural causal model, based on one or more performance data using a learning model with multiple levels. Furthermore, for each of the plurality of new structural causal models, the model updating device 60 generates a plurality of exogenous noise estimates for each of the one or more performance data, and determines whether the causal relationships differ. Then, based on the learning model used for generation and the determination result of whether the causal relationships differ, the model updating device 60 outputs information indicating how the causal relationships of the plurality of variables represented by the one or more performance data have changed. Specifically, the model updating device 60 determines at least one of whether the causal influence has changed, whether the causal structure has changed, and whether the causal order has changed, and outputs the determination result.

[0121] Fig. 9 is a diagram showing an example of information output from the model updating device 60. The model updating device 60 may display the determination results, as to whether or not the causal influence has changed, whether or not the causal structure has changed, and whether or not the causal order has changed, on a display device in the form of text as shown in Fig. 9.

[0122] 10 is a diagram showing an example of the adjacency matrix B' of the updated model. The model updating device 60 may display the updated adjacency matrix B' on the display device in a table format as shown in FIG.

[0123] FIG. 11 is a diagram showing an example of a difference matrix between the adjacency matrix B of the pre-update model and the adjacency matrix B' of the post-update model.

[0124] When updating the structural causal model, the model updating device 60 may generate a difference matrix by calculating the difference between the adjacency matrix B used in the pre-update model and the adjacency matrix B' used in the updated model. Then, the model updating device 60 may output the generated difference matrix by displaying it in a table format on a display device.

[0125] The model updating device 60 may also display an image in which the background of a cell in a table is highlighted according to the magnitude of the numerical value. That is, the display control unit 54 may display the table as a heat map by changing the degree of cell emphasis according to the magnitude of the numerical value in the cell. The model updating device 60 may also change the type of emphasis, such as the color or thickness of a cell, according to the positive or negative change, the magnitude of the absolute value, whether the change is zero or non-zero, the change in the direction of causality, etc.

[0126] FIG. 12 is a diagram showing an example of a bar graph indicating the difference between the adjacency matrix B of the pre-update model and the adjacency matrix B' of the updated model.

[0127] When updating the structural causal model, the model updating device 60 may calculate the difference for each pair of two variables and display the calculated difference in the form of a bar graph as shown in Fig. 12. In this case, the model updating device 60 may sort the pairs of two variables according to the magnitude of the difference and display them in the bar graph.

[0128] FIG. 13 is a diagram showing an example of displaying a causal graph of a pre-update model and a causal graph of a post-update model.

[0129] When the model updating device 60 updates the structural causal model, it may display the causal graph of the pre-update model and the causal graph of the updated model side by side, as shown in Fig. 13. In this case, the model updating device 60 may display the causal graph of the pre-update model and the causal graph of the updated model with the positional relationship of the nodes being the same. In this way, the model updating device 60 allows the difference in the graph structure to be visually recognized.

[0130] FIG. 14 is a diagram showing an example of a displayed causal graph of the updated model.

[0131] When the model updating device 60 updates the structural causal model, it may display a causal graph of the updated model as shown in FIG. 14 . In this case, the model updating device 60 may change the style of the edges, such as the color or thickness of the edges, depending on the difference from the causal graph of the pre-update model. Furthermore, the model updating device 60 may add edges that existed in the pre-update model but do not exist in the updated model, for example, by using dotted lines. Furthermore, the model updating device 60 may add a numerical value to the edge that represents the amount of change from the pre-update model. This allows the model updating device 60 to visually recognize changes in causal influence, causal structure, and causal order. Furthermore, a causal graph of the difference between the pre-update model and the updated model may be displayed.

[0132] As described above, the model updating device 60 according to this embodiment determines whether the causal relationship between the multiple variables sampled from the target system 20 differs from the causal relationship between the multiple variables represented by the structural causal model stored in the model storage device 50, based on the independence between two or more variables in the multiple external noise estimates. For example, the model updating device 60 determines whether the causal relationship between the multiple variables sampled from the target system 20 has changed, based on the independence between two or more variables in the multiple external noise estimates.

[0133] According to such a model updating device 60, it is possible to appropriately update the structural causal model in accordance with changes in the causal relationships of the multiple variables sampled from the target system 20.

[0134] Furthermore, if the model updating device 60 determines that the causal relationships differ, it sequentially generates structural causal models using multiple different learning models, and further determines the generated structural causal models based on the independence between two or more variables in multiple exogenous noise estimates.The model updating device 60 then outputs information indicating which of the causal influence, causal structure, and causal order in the structural causal model has changed, based on the relationship between the learning models and the determination results.

[0135] Conventional devices can generate an adjacency matrix B by re-learning using a causal search method from performance data without detecting changes in causal influence, causal structure, and causal order. However, when the performance data has a small number of samples, a large amount of noise, or noise that is close to a Gaussian distribution, the adjacency matrix B generated by re-learning is unstable and unreliable. It is difficult to determine whether the causal influence, causal structure, and causal order have changed, and to analyze the changes, using an adjacency matrix B with low stability and reliability. However, the model updating device 60 according to this embodiment sequentially generates structural causal models using multiple learning models, and further evaluates each of the generated structural causal models based on the independence between two variables in multiple exogenous noise estimates. This allows for stable and highly reliable detection of changes in the causal influence, causal structure, and causal order of multiple variables sampled from the target system 20.

[0136] Non-Patent Documents 17 and 18 describe techniques for learning the difference between causal models from two data sets. The techniques described in Non-Patent Documents 17 and 18 input two data sets and output the difference between adjacency matrices B. The algorithms of the techniques described in Non-Patent Documents 17 and 18 utilize a technique for testing the difference between precision matrices or regression differences. In contrast, the model updating device 60 according to this embodiment inputs one or more actual data sets and an adjacency matrix B and outputs a determination result of the adjacency matrix B or a change in the adjacency matrix B. The model updating device 60 then uses an algorithm for calculating the independence between two or more variables in multiple external noise estimates. Therefore, the model updating device 60 does not need to input the original data; simply inputting one or more new actual data sets to be analyzed and the pre-update adjacency matrix B can output the updated adjacency matrix B'. The model updating device 60 has the advantage that, when non-Gaussianity is satisfied, the causal order and adjacency matrix B can be estimated from only one or more new actual data sets, i.e., the identifiability of the adjacency matrix B is guaranteed.

[0137] Non-Patent Document 19 describes a technique for learning the difference between causal models from two data sets. The technique described in Non-Patent Document 19 inputs two data sets and outputs the difference between the adjacency matrix B. The algorithm of the technique described in Non-Patent Document 19 utilizes the Hessian of the log-likelihood and conditional independence. The technique described in Non-Patent Document 19 also guarantees identifiability when non-linearity is satisfied. In contrast, the model updating device 60 according to this embodiment has the advantage that it does not need to input the original data, but can output the updated adjacency matrix B' by simply inputting one or more new performance data sets to be analyzed and the pre-update adjacency matrix B. The model updating device 60 also has the advantage that the identifiability of the adjacency matrix B is guaranteed when non-Gaussianity is satisfied.

[0138] 15 is a diagram showing a modified example of the information processing system 10 according to the embodiment. The information processing system 10 according to the modified example further includes an analysis device 30 and an abnormal change detection device 80.

[0139] The model updating device 60 periodically acquires one or more pieces of performance data including performance values ​​of multiple variables, determines whether the data contradicts the causal relationships, and performs a series of processes to update the structural causal model if it determines that the causal relationships contradict each other, for example, as a routine operation. Instead of or in addition to such processes, the model updating device 60 according to the modified example performs a series of processes as a non-routine operation in response to instructions from the abnormal change detection device 80.

[0140] The analysis device 30 is an information processing device that executes information processing. The analysis device 30 acquires one or more pieces of performance data from the target system 20, each of which includes a plurality of performance values ​​corresponding to a plurality of variables. Prior to processing, the analysis device 30 is provided with a structural causal model stored in the model storage device 50, and the structural causal model is set. Based on the acquired one or more pieces of performance data and the preset structural causal model, the analysis device 30 generates cause information that indicates the cause of a change in at least one variable among the plurality of variables. The analysis device 30 then outputs the cause information, for example, by displaying it.

[0141] In this embodiment, the analysis device 30 acquires n samples of performance data (n is an integer equal to or greater than 1). In this embodiment, each of the one or more performance data is assigned an index that identifies the conditions, such as the time at which it was sampled.

[0142] Such an analysis device 30 can output causal information indicating the direct and indirect causes of changes in variables by using a structural causal model.

[0143] The abnormal change detection device 80 acquires one or more pieces of performance data, etc., and detects that an abnormality or a possible abnormality has occurred in the target system 20, or that the state of the target system 20 has changed. When the abnormal change detection device 80 detects a change in the state of the target system 20, it issues an operation instruction to the model updating device 60. For example, the abnormal change detection device 80 may determine the state of the target system 20 by comparing the quantile, standard deviation, etc. of the actual value of one of the multiple variables to be monitored with a preset threshold. Furthermore, for example, the abnormal change detection device 80 may determine the state of the target system 20 by testing the mean, standard deviation, etc. of the actual value of the monitored variable. Furthermore, for example, the abnormal change detection device 80 may individually determine each of two or more monitored variables among the multiple variables, and determine that an abnormality has occurred, changed, or both an abnormality and change in the state of the target system 20 when an abnormality is detected in one or a predetermined number of the two or more monitored variables.

[0144] Furthermore, the abnormal change detection device 80 may perform multivariate analysis on two or more monitored variables among the plurality of variables to determine the state of the target system 20. For example, the abnormal change detection device 80 may determine the state of the target system 20 using a method based on Hotelling, k-NN, SVM, CAE, a graphical model, a density ratio, or the like for two or more monitored variables.

[0145] By being equipped with such an abnormal change detection device 80, the information processing system 10 can assist in the task of identifying the abnormalities, changes, or factors that caused the abnormalities and changes detected by the abnormal change detection device 80.

[0146] Furthermore, when the target system 20 is a system for manufacturing a product, the target system 20 uses a control chart to monitor and control the quality characteristics of the product. The control chart includes information that defines the control range or specification range of values ​​that represent the quality characteristics of the product.

[0147] In this case, the multiple variables include the quality characteristics of the products defined in the control chart as variables, and the structural causal model updated by the model updating device 60 includes the quality characteristics of the products defined in the control chart as variables.

[0148] The abnormal change detection device 80 causes the model updating device 60 to execute processing when the actual value of a variable representing the quality characteristics of a product falls outside the control range or specification range. When the quality characteristics of a product fall outside the control range or specification range shown on the control chart, the model updating device 60 acquires one or more pieces of actual data including the actual values ​​of multiple variables, determines whether the data deviates from the causal relationship, and, if it determines that the causal relationship deviates, performs a series of processes to update the structural causal model. This allows the user to combine the structural causal model updated by the model updating device 60, information indicating the content of the change, and the control chart to investigate the cause and consider countermeasures.

[0149] (Analyzer 30) FIG. 16 is a diagram showing an example of the configuration of the analysis device 30. As shown in FIG.

[0150] For example, the analysis device 30 may have a configuration as shown in Fig. 16. The analysis device 30 includes an actual data acquisition unit 32, an actual data storage unit 34, a model acquisition unit 36, a model storage unit 38, an external noise estimation unit 40, a contribution decomposition unit 42, a decomposition result storage unit 44, and an output unit 46.

[0151] The performance data acquiring unit 32 acquires one or more performance data from the target system 20. In this embodiment, the performance data acquiring unit 32 acquires performance data including d performance values ​​corresponding to d variables, for n samples.

[0152] The performance data storage unit 34 stores one or more acquired performance data. In this embodiment, the performance data storage unit 34 stores a performance data matrix X'.

[0153] The model acquisition unit 36 ​​acquires the structural causal model from the model storage device 50. In this embodiment, the model acquisition unit 36 ​​acquires the adjacency matrix B' from the model storage device 50.

[0154] The model storage unit 38 stores the structural causal model acquired by the model acquisition unit 36. In this embodiment, the model storage unit 38 stores the adjacency matrix B'.

[0155] The exogenous noise estimation unit 40 calculates a plurality of exogenous noise estimated values ​​corresponding to a plurality of variables for each of the one or more pieces of actual data stored in the actual data storage unit 34 and the structural causal model acquired by the model acquisition unit 36. Note that the exogenous noise estimation unit 40 may use the exogenous noise estimated value estimated by the estimation unit 60.

[0156] In this embodiment, the external noise estimation unit 40 calculates the external noise matrix E' by equation (8).

number

[0157] Based on the structural causal model and multiple exogenous noise estimates for each of the one or more actual data, the contribution decomposition unit 42 generates, for each combination of two variables included in the multiple variables, a contribution that represents the magnitude of the influence that exogenous noise given to a source variable, one of the two variables, has on a target variable, the other of the two variables, for each of the one or more actual data.

[0158] The contribution will be further explained below.

[0159] By rearranging equation (8), the performance data matrix X' is expressed as in equation (9).

number

[0160] (IB´) -1 When is defined as the coefficient matrix A, the performance data matrix X' is expressed as shown in equation (10).

number

[0161] If the causal graph is an acyclic graph, by appropriately rearranging the order of the variables, the adjacency matrix B becomes a strictly upper triangular matrix with zero diagonal elements. Therefore, (IB') has an inverse matrix. That is, the coefficient matrix A = (IB'), which is the inverse matrix of (IB). -1 is an upper triangular matrix with diagonal elements of 1. The order of multiple variables when arranged in this way is called the causal order. A causal graph representing a structural causal model may contain a directed edge from a node corresponding to any first-order variable to a node corresponding to any second-order variable whose causal order is greater than first. However, the causal graph does not contain a directed edge from a node corresponding to a second-order variable to a node corresponding to the first-order variable.

[0162] Also, if the causal graph is not acyclic, the inverse matrix of (IB') does not necessarily exist. Therefore, the coefficient matrix A is (IB') -1 Instead, we use the generalized inverse matrix (IB´) + may be.

[0163] The element in row i and column k (k is an integer between 1 and d) in X' = E'A is x ik x ik is expressed as in equation (11).

number

[0164] x represented by equation (11) ik represents the actual value of the variable with index k out of d variables that is included in the actual data of the sample with index i out of n samples.

[0165] The right side of equation (11) is a linear sum of multiple terms. ik is expressed as a linear sum equation that sums up multiple terms.

[0166] The linear sum equation on the right side of equation (11) is (k-1) A jk e ij and one e ik This includes the terms

[0167] e ik represents the value of the element in row i and column k in the extrinsic noise matrix E'. ik is an exogenous noise estimate representing the exogenous noise given to the variable with index k in the historical data of the sample with index i.

[0168] e ij represents the value of the element in the i-th row and j-th column of the extrinsic noise matrix E'. ijis an exogenous noise estimate representing the exogenous noise given to the variable with index j in the historical data of the sample with index i.

[0169] A jk is the coefficient matrix A=(IB´) -1 represents the value of the element in row j and column k in

[0170] Here, in equation (11), the variable with index j is located upstream in the causal graph from the variable with index k. In other words, the variable with index j has a lower causal order than the variable with index k. Therefore, e ij is the actual value of the variable with index k in the sample with index i, x ik The external noise estimate represents the external noise given to a variable that may directly affect the variable, or the external noise estimate represents the external noise given to a variable that may indirectly affect the variable.

[0171] In other words, A in the linear sum equation on the right side of equation (11) jk e ij is the actual value of the variable with index k in the sample with index i, x ik It represents the influence of exogenous noise on variables that may directly or indirectly affect the

[0172] If the adjacency matrix B' is a strict upper triangular matrix with diagonal elements set to zero, the coefficient matrix A will be an upper triangular matrix with diagonal elements set to 1 and elements below the diagonal elements set to 0. Therefore, in this case, equation (11) may be expressed as equation (12).

number

[0173] The linear sum expression on the right side of equation (12) is the multiple exogenous noise estimates e for the actual data of sample with index i. ij and the corresponding coefficients in the coefficient matrix A.jk That is, the linear sum equation on the right side of equation (12) is an equation that performs a product-sum operation on the row corresponding to the actual data of the sample with index i in the extrinsic noise matrix E' and the column corresponding to the variable with index k in the coefficient matrix A. However, since the coefficient matrix A is an upper triangular matrix with 0 below the diagonal elements, A in the linear sum equation on the right side of equation (12) jk e ij is 0 if j is greater than k.

[0174] The contribution decomposition unit 42 calculates A jk e ij is generated as a contribution for each combination of two variables included in the plurality of variables for each of one or more performance data. That is, the contribution decomposition unit 42 generates A jk e ij Let x be the actual value of the variable with index k contained in the sample with index i. ik for the variable with index k contained in the sample with index i. ik The contribution decomposition unit 42 generates the contribution as the degree of contribution of external noise given to a source variable, which is one of two variables, on a target variable, which is the other of the two variables. Note that the contribution decomposition unit 42 also generates the contribution for a combination of two variables where the source variable and the target variable are the same variable.

[0175] For example, when a first variable among a plurality of variables in a first piece of one or more pieces of actual data is set as a target variable and a second variable among the plurality of variables is set as a source variable, the contribution to the combination of the two variables is the value of a term among a plurality of terms included in a linear sum formula, the term including an extrinsic noise estimated value representing the extrinsic noise given to the second variable. In this case, the linear sum formula is a formula that performs a product-sum operation on the row corresponding to the first actual data in the extrinsic noise matrix E' and the column corresponding to the first variable in the coefficient matrix A. Furthermore, the coefficient matrix A is the inverse matrix (IB') of the matrix (IB') obtained by subtracting the adjacent matrix B' from the unit matrix I. -1 or generalized inverse matrix (IB´) + is.

[0176] For example, for the actual data of the sample with index i (first actual data), if the variable with index k (first variable) is the target variable and the variable with index j (second variable) is the source variable, the contribution is A jk e ij For the actual data of the sample with index i (first actual data), if the target variable and source variable are both variables with index k (first variable), the contribution is e ik This becomes:

[0177] In this embodiment, the contribution decomposition unit 42 generates a contribution tensor C, which is a third-order tensor. The contribution tensor C includes an i-component, a j-component, and a k-component.

[0178] The i component corresponds to each of one or more pieces of performance data. In this embodiment, the i component corresponds to each of the performance data of n samples.

[0179] The j component and the k component correspond to each of a plurality of variables. In this embodiment, the j component and the k component correspond to each of the d variables.

[0180] In this embodiment, the contribution decomposition unit 42 decomposes C into i, j, and k components of the contribution tensor C. ijkis calculated using equation (13).

number

[0181] That is, each of the multiple elements included in the contribution tensor C represents the contribution of one or more pieces of actual data identified by the i component, with the variable identified by the j component of the multiple variables being the source variable and the variable identified by the k component of the multiple variables being the target variable.

[0182] For the multiple elements contained in such a contribution tensor C, equation (14) holds.

number

[0183] That is, each of the multiple elements included in the contribution tensor C represents a term included in the linear sum equation for each of one or more pieces of performance data. Therefore, by calculating such a contribution tensor C, the contribution decomposition unit 42 can calculate the results of a comprehensive analysis of the presence of other variables that cause a change in at least one variable among the multiple variables.

[0184] The contribution decomposition unit 42 may include values ​​obtained by converting the numerical value shown in formula (13) according to a predetermined rule in the elements of the contribution tensor C. For example, the contribution decomposition unit 42 may include a sign indicating whether the numerical value shown in formula (13) is positive or negative, an absolute value, a value indicating whether the numerical value is included in a range from a predetermined upper limit value to a predetermined lower limit value, or a value obtained by leveling the numerical value shown in formula (13) into a predetermined level.

[0185] The decomposition result storage unit 44 stores the contribution degree for each combination of two variables included in the plurality of variables for each of one or more pieces of performance data calculated by the contribution decomposition unit 42. In this embodiment, the decomposition result storage unit 44 stores a contribution tensor C.

[0186] The output unit 46 outputs, for example, by displaying on a display device, the contribution degree for each pair of two variables included in the plurality of variables for each of one or more pieces of performance data stored in the decomposition result storage unit 44. In this embodiment, the output unit 46 outputs the value of at least one element included in the contribution tensor C stored in the decomposition result storage unit 44, or a value obtained by converting the value of at least one element included in the contribution tensor C stored in the decomposition result storage unit 44.

[0187] The output unit 46 may accept a selection of at least one piece of performance data of interest indicating a sample of interest from among the one or more pieces of performance data. The output unit 46 may also accept a selection of at least one source variable of interest to be selected as a source variable from among the plurality of variables. The output unit 46 may also accept a selection of at least one target variable of interest to be selected as a target variable from among the plurality of variables. The output unit 46 may accept the selection of the at least one piece of performance data of interest, the at least one source variable of interest, and the at least one target variable of interest in response to a user operation, or may accept the selection based on preset information.

[0188] Then, the output unit 46 outputs, as cause information, contribution degrees corresponding to all or some combinations of the selected at least one piece of note performance data, the selected at least one source variable of interest, and the selected target variable of interest. For example, the output unit 46 causes a display device or the like to display cause information in the form of a table or graph representing contribution degrees corresponding to all or some combinations of the selected at least one piece of note performance data, the selected source variable of interest, and the selected target variable of interest.

[0189] (Hardware configuration of the model update device 60, etc.) Fig. 17 is a diagram showing an example of the hardware configuration of the model updating device 60. The model updating device 60 is realized by, for example, an information processing device having the hardware configuration shown in Fig. 17. The model updating device 60 includes a CPU (Central Processing Unit) 901, a RAM (Random Access Memory) 902, a ROM (Read Only Memory) 903, a storage device 904, and a communication interface device 905. These components are connected via a bus.

[0190] The CPU 901 is one or more processors that execute arithmetic processing, control processing, etc. according to a program. The CPU 901 uses a predetermined area of ​​the RAM 902 as a working area and executes various processes in cooperation with programs stored in the ROM 903, the storage device 904, etc.

[0191] The RAM 902 is a memory such as an SDRAM (Synchronous Dynamic Random Access Memory), and functions as a work area for the CPU 901. The ROM 903 is a memory that stores programs and various information in a non-rewritable manner.

[0192] The storage device 904 is a device that writes and reads data to a semiconductor storage medium such as a flash memory, or a magnetically or optically recordable storage medium, etc. The storage device 904 writes and reads data to the storage medium in response to control from the CPU 901. The communication interface device 905 communicates with external devices via a network in response to control from the CPU 901.

[0193] A program executed by the information processing device causes the information processing device to function as the model update device 60. This program is loaded on the RAM 902 by the CPU 901 (processor) and executed.

[0194] In addition, the program executed by the information processing device is provided as a file in a format that can be installed on the information processing device or in an executable format, recorded on a recording medium that can be read by the information processing device, such as a CD-ROM, flexible disk, CD-R, or DVD (Digital Versatile Disk).

[0195] This program may also be stored on a computer connected to a network such as the Internet and provided by being downloaded via the network. This program may also be provided or distributed via a network such as the Internet. The program executed by the model updating device 60 may also be provided by being pre-installed in the ROM 903 or the like.

[0196] The program for causing the information processing device to function as the model updating device 60 includes, for example, a data acquisition module, an estimation module, a determination module, a learning module, a control module, a result output module, and an update module. When this program is executed by the CPU 901, each module is loaded into the RAM 902, causing the CPU 901 to function as the data acquisition unit 62, the estimation unit 66, the determination unit 68, the learning unit 70, the control unit 74, the result output unit 76, and the update unit 78. If the CPU 901 has multiple processors, these units may be divided among the multiple processors. Note that these components may be partially or entirely implemented by hardware. Furthermore, this program causes the RAM 902 and the storage device 904 to function as the data storage unit 64 and the updated model storage unit 72.

[0197] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, and are also included in the scope of the invention and its equivalents as defined in the claims. [Explanation of symbols]

[0198] 10 Information Processing Systems 20 Target System 30 Analyzer 50 Model Storage Devices 60 Model Update Device 62 Data Acquisition Section 64 Data storage unit 66 Estimation part 68 Judgment section 70 Learning Department 72 Update model storage unit 74 Control Unit 76 Result output section 78 Update section 80 Abnormal Change Detection Device

Claims

1. generating, for each of the one or more pieces of performance data, a plurality of external noise estimates corresponding to the plurality of variables based on one or more pieces of performance data each including a plurality of performance values ​​corresponding to a plurality of variables and a pre-update model which is a structural causal model representing a causal relationship between the plurality of variables, each of the plurality of external noise estimates representing an estimate of an influence of external noise different from an influence from the plurality of variables on a corresponding variable among the plurality of variables; determining whether or not a causal relationship between the plurality of variables represented by the one or more pieces of performance data differs from a causal relationship between the plurality of variables represented by the pre-update model, based on independence between any two or more variables in the plurality of external noise estimates for each of the one or more pieces of performance data; If it is determined that the causal relationships are different, an updated model that is the structural causal model is generated based on the one or more pieces of performance data. Processing section An information processing device comprising:

2. the structural causal model is expressed using an adjacency matrix that indicates, for each combination of two variables included in the plurality of variables, the magnitude of the influence from one variable to the other variable; The processing unit an external noise matrix is ​​calculated by multiplying an actual data matrix including the one or more actual data by a matrix obtained by subtracting the adjacent matrix from a unit matrix, the external noise matrix including the plurality of external noise estimation values ​​for each of the one or more actual data; Based on the extrinsic noise matrix, it is determined whether the causal relationships of the plurality of variables represented by the one or more pieces of performance data differ from the causal relationships of the plurality of variables represented by the pre-update model. The information processing device according to claim 1 .

3. The processing unit For each combination of two variables included in the plurality of variables, calculate an independence index value that indicates independence between two or more columns or rows corresponding to the two variables in the extrinsic noise matrix; comparing the calculated independence index value with a threshold for each combination of the two variables included in the plurality of variables; Based on a comparison result between the independence index value for each combination of the two variables included in the plurality of variables and the threshold, it is determined whether or not the causal relationship of the plurality of variables represented by the one or more pieces of performance data differs from the causal relationship of the plurality of variables represented by the pre-update model. The information processing device according to claim 2 .

4. The processing unit For each combination of two variables included in the plurality of variables, calculate an independence index value that indicates independence between two or more columns or rows corresponding to the two variables in the extrinsic noise matrix; calculating a statistical value by integrating the calculated independence index values ​​for each combination of the two variables included in the plurality of variables; Based on a comparison result between the statistical value and a threshold, it is determined whether or not the causal relationship of the plurality of variables represented by the one or more pieces of performance data differs from the causal relationship of the plurality of variables represented by the pre-update model. The information processing device according to claim 2 .

5. The processing unit calculates, as the independence index value, any one of a correlation coefficient, a rank correlation coefficient, a mutual information amount, a pairwise likelihood ratio, a distance correlation, a Hilbert-Schmidt Independence Criterion (HSIC), a Maximal Information Coefficient (MIC), and a Randomized Dependence Coefficient (RDC) between two columns corresponding to the two variables in the extrinsic noise matrix, or a value obtained by combining two or more of these. The information processing device according to claim 3 .

6. the structural causal model is expressed using an adjacency matrix that represents, for each combination of two variables included in the plurality of variables, an influence from one variable to the other variable; The processing unit If it is determined that the causal relationships are different, new values ​​corresponding to non-zero elements in the adjacency matrix of the pre-update model are estimated using linear regression based on the one or more pieces of actual data; The updated model is generated by updating the values ​​of non-zero elements in the adjacency matrix of the pre-update model to the estimated new values. The information processing device according to claim 1 .

7. the structural causal model is expressed using an adjacency matrix that indicates, for each combination of two variables included in the plurality of variables, the magnitude of the influence from one variable to the other variable; When it is determined that the causal relationships are different, the processing unit sets the causal order to be the same as the adjacency matrix of the pre-update model, and generates the updated model using regularized linear regression based on the one or more pieces of performance data. The information processing device according to claim 1 .

8. When it is determined that the causal relationships are different, the processing unit generates the updated model using Lasso or Adaptive Lasso based on the one or more pieces of performance data so that the causal order is the same as the adjacency matrix of the pre-update model. The information processing device according to claim 7 .

9. When it is determined that the causal relationships are different, the processing unit generates the updated model using one of Adaptive Lasso, Transfer Lasso, or Adaptive Transfer Lasso, with the pre-update model as an initial estimate based on the one or more pieces of performance data, so that the causal order is the same as the adjacency matrix of the pre-update model. The information processing device according to claim 7 .

10. When it is determined that the causal relationships are different, the processing unit generates the updated model including the causal order based on the one or more pieces of performance data. The information processing device according to claim 1 .

11. The processing unit generates the updated model based on a causal search method based on non-Gaussianity. The information processing device according to claim 10.

12. The processing unit generates the updated model based on ICA-LiNGAM or DirectLiNGAM. The information processing device according to claim 11.

13. the structural causal model is expressed using an adjacency matrix that indicates, for each combination of two variables included in the plurality of variables, the magnitude of the influence from one variable to the other variable; The processing unit generating a plurality of first external noise estimated values, which are the plurality of external noise estimated values ​​for each of the one or more pieces of actual data, based on the one or more pieces of actual data and the pre-update model; performing a first determination process to determine whether or not a causal relationship between the plurality of variables represented by the one or more pieces of performance data differs from a causal relationship between the plurality of variables represented by the pre-update model, based on independence between any two or more variables in the plurality of first exogenous noise estimates for each of the one or more pieces of performance data; When it is determined that there is no difference in the causal relationship by the first determination process, outputting at least one of information indicating that there is no change in the causal influence, information indicating that there is no change in the causal structure, and information indicating that there is no change in the causal order; if it is determined by the first determination process that the causal relationships differ, estimating new values ​​corresponding to non-zero elements in the adjacency matrix of the pre-update model using linear regression based on the one or more pieces of performance data, and updating the values ​​of the non-zero elements in the adjacency matrix of the pre-update model to the estimated new values, thereby generating a first structural causal model that is the structural causal model; generating a plurality of second exogenous noise estimates, which are the plurality of exogenous noise estimates for each of the one or more pieces of performance data, based on the one or more pieces of performance data and the first structural causal model; performing a second determination process to determine whether or not a causal relationship between the plurality of variables represented by the one or more pieces of performance data differs from a causal relationship between the plurality of variables represented by the first structural causal model, based on the independence between any two or more variables in the plurality of second external noise estimates for each of the one or more pieces of performance data; When it is determined by the second determination process that there is no difference in the causal relationship, at least one of information indicating that there is a change in the causal influence, that there is no change in the causal structure, and that there is no change in the causal order is output. The information processing device according to claim 1 .

14. The processing unit when it is determined by the second determination process that the causal relationships are different, a causal order is made the same as the adjacency matrix of the pre-update model, and a second structural causal model is generated as the structural causal model using regularized linear regression based on the one or more pieces of performance data; generating a plurality of third exogenous noise estimates, which are the plurality of exogenous noise estimates for each of the one or more pieces of performance data, based on the one or more pieces of performance data and the second structural causal model; performing a third determination process to determine whether or not a causal relationship between the plurality of variables represented by the one or more pieces of performance data differs from a causal relationship between the plurality of variables represented by the second structural causal model, based on the independence between any two or more variables in the plurality of third external noise estimates for each of the one or more pieces of performance data; When it is determined that there is no difference in the causal relationship by the third determination process, outputting at least one of information indicating that there is a change in the causal influence, information indicating that there is a change in the causal structure, and information indicating that there is no change in the causal order; When it is determined that the causal relationships are different by the third determination process, at least one of information indicating that there is a change in causal influence, a change in causal structure, and a change in causal order is output. The information processing device according to claim 13.

15. the structural causal model is expressed using an adjacency matrix that indicates, for each combination of two variables included in the plurality of variables, the magnitude of the influence from one variable to the other variable; The processing unit generating a plurality of first external noise estimated values, which are the plurality of external noise estimated values ​​for each of the one or more pieces of actual data, based on the one or more pieces of actual data and the pre-update model; performing a first determination process to determine whether or not a causal relationship between the plurality of variables represented by the one or more pieces of performance data differs from a causal relationship between the plurality of variables represented by the pre-update model, based on independence between any two or more variables in the plurality of first exogenous noise estimates for each of the one or more pieces of performance data; When it is determined that there is no difference in the causal relationship by the first determination process, outputting at least one of information indicating that there is no causal influence, information indicating that there is no change in the causal structure, and information indicating that there is no change in the causal order; if it is determined by the first determination process that the causal relationships are different, a causal order is made the same as the adjacency matrix of the pre-update model, and a second structural causal model is generated as the structural causal model using regularized linear regression based on the one or more pieces of performance data; generating a plurality of third exogenous noise estimates, which are the plurality of exogenous noise estimates for each of the one or more pieces of performance data, based on the one or more pieces of performance data and the second structural causal model; performing a third determination process to determine whether or not a causal relationship between the plurality of variables represented by the one or more pieces of performance data differs from a causal relationship between the plurality of variables represented by the second structural causal model, based on the independence between any two or more variables in the plurality of third external noise estimates for each of the one or more pieces of performance data; When it is determined that there is no difference in the causal relationship by the third determination process, at least one of information indicating that there is a change in the causal influence, information indicating that there is a change in the causal structure, and information indicating that there is no change in the causal order is output; When it is determined that the causal relationships are different by the third determination process, at least one of information indicating that there is a change in causal influence, a change in causal structure, and a change in causal order is output. The information processing device according to claim 1 .

16. The processing unit generating the plurality of exogenous noise estimates for each of the one or more pieces of historical data based on the one or more pieces of historical data and the updated model; Based on the independence between any two or more variables in the plurality of exogenous noise estimates after updating for each of the one or more performance data, at least one of information indicating whether or not there is a change in causal influence, whether or not there is a change in the causal structure, and whether or not there is a change in the causal order is output. The information processing device according to claim 1 .

17. the structural causal model is expressed using an adjacency matrix that indicates, for each combination of two variables included in the plurality of variables, the magnitude of the influence from one variable to the other variable; The processing unit outputs the adjacency matrix used in the updated model, or a difference matrix between the adjacency matrix used in the pre-updated model and the adjacency matrix used in the updated model. The information processing device according to claim 1 .

18. the structural causal model is expressed using an adjacency matrix that indicates, for each combination of two variables included in the plurality of variables, the magnitude of the influence from one variable to the other variable; The processing unit displays a causal model representing the adjacency matrix used in the updated structural causal model. The information processing device according to claim 1 .

19. An information processing device according to any one of claims 1 to 18; an analytical device; Equipped with The analysis device calculating the plurality of external noise estimates based on the one or more pieces of performance data and the structural causal model; Based on the structural causal model and the plurality of exogenous noise estimates for each of the one or more pieces of actual data, a contribution is generated for each combination of two variables included in the plurality of variables, the contribution representing the magnitude of the influence that the exogenous noise imparted to a source variable, which is one of the two variables, has on a target variable, which is the other of the two variables. Information processing system.

20. An information processing device according to any one of claims 1 to 18; a change detection device that detects a change in the state of the target system that outputs the one or more pieces of performance data; Equipped with When the change detection device detects a change in the state of the target system, the change detection device causes the information processing device to execute a process. Information processing system.

21. the target system is a system for manufacturing a product, the plurality of variables includes a quality characteristic of the product as a variable; The change detection device causes the information processing device to execute processing when an actual value of a variable representing a quality characteristic of the product falls outside a control range or a specification range defined by a control chart generated in advance.

21. The information processing system according to claim 20.

22. An information processing method executed by an information processing device, the information processing device generates, for each of the one or more pieces of performance data, a plurality of external noise estimated values ​​corresponding to the plurality of variables, based on one or more pieces of performance data including a plurality of performance values ​​each corresponding to a plurality of variables, and a pre-update model that is a structural causal model representing a causal relationship between the plurality of variables, and each of the plurality of external noise estimated values ​​represents an estimated value of an influence of external noise different from an influence from the plurality of variables on a corresponding variable among the plurality of variables; the information processing device determines, based on independence between any two or more variables in the plurality of external noise estimation values ​​for each of the one or more performance data, whether or not a causal relationship between the plurality of variables represented by the one or more performance data differs from a causal relationship between the plurality of variables represented by the pre-update model; When the information processing device determines that the causal relationships are different, it generates an updated model that is the structural causal model based on the one or more pieces of performance data. Information processing methods.

23. Computer, generating, for each of the one or more pieces of performance data, a plurality of external noise estimates corresponding to the plurality of variables based on one or more pieces of performance data each including a plurality of performance values ​​corresponding to a plurality of variables and a pre-update model which is a structural causal model representing a causal relationship between the plurality of variables, each of the plurality of external noise estimates representing an estimate of an influence of external noise different from an influence from the plurality of variables on a corresponding variable among the plurality of variables; determining whether or not a causal relationship between the plurality of variables represented by the one or more pieces of performance data differs from a causal relationship between the plurality of variables represented by the pre-update model, based on independence between any two or more variables in the plurality of external noise estimates for each of the one or more pieces of performance data; If it is determined that the causal relationships are different, an updated model that is the structural causal model is generated based on the one or more pieces of performance data. Processing section A program to function as a

Images