Information processing with an objective function that includes a generated regularization term using an apparatus, information processing method, and non-transitory computer readable medium
An information processing apparatus according to an embodiment of the present invention includes a group setting device configured to group a plurality of variables included in operation data into at least one type of groups, based on structure data representing a structural relation among the plurality of variables; a regularization term generator configured to generate at least one regularization term corresponding to the at least one type based on a coefficient for the variable included in the at least one type of groups, and a coefficient estimator configured to estimate, based on the operation data and an objective function including the at least one regularization term, values of a plurality of the coefficients for the plurality of variables.
1 . An information processing apparatus comprising:
processing circuitry configured to:
read operation data of a target apparatus including a plurality of variables obtained by sensing a plurality of components in the target apparatus with a plurality of sensors monitoring the plurality of components and structure data representing structural relation among the plurality of variables from at least one memory, the operation data including whether or not a predetermined event has occurred,
group variables having a same type of structural relation among the plurality of variables in the operating data into a same group based on the structure data to generate a plurality of groups of at least one group type corresponding to a type of structural relation of variables in the groups, each of the plurality of variables being included in one of the plurality of groups,
determine at least one regularization type corresponding to the plurality of groups based on mapping data mapping a plurality of group types to a plurality of regularization types and based on group types of the plurality of groups, and generate at least one regularization term including a coefficient for the variables included in the groups for the at least one regularization type, select an evaluation function to be used among a plurality of evaluation functions, depending on whether or not there is an overlap of variables among the plurality of groups, each evaluation function including the coefficients for the plurality of variables and the plurality of variables,
generate an objective function including the least one regularization term generated for the at least one regularization type and including the selected evaluation function,
estimate, based on the operation data and the objective function, values of a plurality of coefficients for the plurality of variables,
select only variables with an absolute value greater than zero from the plurality of variables based on the values of the estimated coefficients, and
identify, as drawing variables, the selected variables and their structurally adjacent variables determined based on the structure data,
generate, in a form of a table or graph, variable structure data that visually represents the drawing variables and structural relationships among the drawing variables, wherein the selected variables correspond to components having more influence on occurrence of the predetermined event among the plurality of components in the target apparatus and the non-selected variables correspond to components less influence on occurrence of the predetermined event among the plurality of components; and
a display circuit configured to display the variable structure data visibly to a user,
wherein the processing circuitry constructs, based on a value of the coefficient estimated for the selected variable and the selected variable, a model of detecting occurrence of the predetermined event in the target apparatus, and
the apparatus comprises receiving circuitry configured to receive an operation data acquired by the plurality of sensors via a network, and
the processing circuitry calculates an output value of the model based on the operation data received in the receiving circuitry, and detects the occurrence of the predetermined event in the target apparatus based on the output value, wherein
the display displays the variable structure data when the processing circuitry detects the occurrence of the predetermined event in the target apparatus so that the user is able to specify a component that caused the occurrence of the predetermined event among the plurality of components in the target apparatus to one of the components corresponding to the selected variables,
the operation data includes first operation data acquired in a first period and second operation data acquired in a second period later than the first period,
the processing circuitry is configured to perform processing on the first operation data to select first variables being the variables with the absolute value greater than zero from the plurality of variables, and perform processing on the second operation data to select second variables being the variables with the absolute value greater than zero from the plurality of variables,
the processing circuitry is configured to determine whether a first set of drawing corresponding to the first operation data structurally differs a second set of drawing variables corresponding to the second operation data and, when a structural difference is detected between the first and second sets of drawing variables, output an alert indicating that mechanism of occurrence of the predetermined event in the target apparatus has changed.
2 . The apparatus according to claim 1 , wherein the at least one group type indicates a group of a class relation integrating variables corresponding to instances of a class, there are two or more groups each having the class relationship, there is no overlap of variables among the groups, the processing circuitry is configured to generate a regularization term of Group Lasso for the groups of the class relation, and the drawing variables include both selected variables and adjacent variables, each corresponding to an instance of the class associated with its group.
3 . The apparatus according to claim 1 , wherein the at least one group type indicates a group of a hierarchical relation integrating variables corresponding to low-order nodes having a common high-order node in a tree structure, there are two or more groups each having the hierarchical relation, there is no overlap of variables among the groups, the processing circuitry is configured to generate a regularization term of Overlapping Group Lasso for the groups of the hierarchical relation, and the drawing variables include both selected variables corresponding to the low-order nodes and structurally adjacent variables including the high-order node associated with the hierarchical relation.
4 . The apparatus according to claim 3 , wherein the high-order node is a parent node of the plurality of low-order nodes and the drawing variables include at least one of the selected variables corresponding to the low-order nodes and the parent node as a structurally adjacent variable.
5 . The apparatus according to claim 1 , wherein the at least one group type indicates a group of a reference relation integrating variables having a reference relation, the processing circuitry is configured to generate a regularization term of Clustered Lasso for the group of the reference relation, and the drawing variables include selected variables belonging to the group of the reference relation and adjacent variables connected via the reference relation.
6 . The apparatus according to claim 1 , further comprising:
a regularization-rule input circuitry to receive a value of a regularization parameter included in the regularization term, wherein
the processing circuitry is configured to generate the regularization term using the input value of the regularization parameter to generate variable structure data including drawing variables based on the generated regularization term.
7 . The apparatus according to claim 1 , comprising:
a display to display the variable structure data, wherein
a regularization-rule input circuitry receives a plurality of values of the regularization term,
the processing circuitry is configured to generate the variable structure data for each of the plurality of values of the regularization parameter,
the display comparably displays a plurality of sets of the plurality of values and a plurality of the variable structure data corresponding to the plurality of values, and
each of the variable structure data includes drawing variables representing selected variables and adjacent variables based on a structure relation.
8 . The apparatus according to claim 1 , wherein the processing circuitry sums the at least one regularization term generated for the at least one regularization type and the selected evaluation function to generate the objective function, or subtracts the at least one regularization term generated for the at least one regularization type from the selected evaluation function to generate the objective function, and the objective function is used to identify selected variables and generate drawing variables including their adjacent variables.
9 . The apparatus according to claim 8 , wherein the processing circuitry minimizes or maximizes the objective function to obtain the values of a plurality of coefficients for the plurality of variables and generates variable structure data including drawing variables based on the result of the optimization.
10 . The apparatus according to claim 1 , wherein
the processing circuitry optimizes the objective function based on, the operation data to estimate the values of the plurality of coefficients for the plurality of variables and extracts a function part of the optimized objective function, which includes the selected variables and the coefficients of the selected variables, and which does not include non-selected variables and the coefficients of the non-selected variables, wherein the extracted function part corresponds to the model of detecting occurrence of the predetermined event in the target apparatus, and the drawing variables are based on the selected variables used in the model and their structurally adjacent variables.
11 . The apparatus according to claim 1 , wherein the predetermined event is abnormality, and the drawing variables represent components related to occurrence of the abnormality, including both selected variables and structurally adjacent variables.
12 . The apparatus according to claim 1 , wherein the at least one group type indicates a plurality of group types,
the plurality of group types indicates a group of a class relation integrating variables corresponding to instances of a class, and a group of a hierarchical relation integrating variables corresponding to low-order nodes having a common high-order node in a tree structure,
there are two or more groups each having the class relationship, there is no overlap of variables among the groups, and the processing circuitry is configured to generate a regularization term of Group Lasso for the groups of the class relation,
there are two or more groups each having the hierarchical relation, there is no overlap of variables among the groups, and the processing circuitry is configured to generate a regularization term of Overlapping Group Lasso for the groups of the hierarchical relation, and
the drawing variables include selected variables and adjacent variables derived from each of the class relation and the hierarchical relation.
13 . An information processing method performed by a computer including a processor and a memory comprising:
reading operation data of a target apparatus including a plurality of variables obtained by sensing a plurality of components in the target apparatus with a plurality of sensors monitoring the plurality of components in the target apparatus and structure data representing structural relation among the plurality of variables from at least one memory, the operation data including whether or not a predetermined event has occurred;
grouping variables having a same type of structural relation among the plurality of variables in the operating data into a same group based on the structure data to generate a plurality of groups of at least one group type corresponding to a type of structural relation of variables in the groups, each of the plurality of variables being included in one of the plurality of groups;
determining at least one regularization type corresponding to the plurality of groups based on mapping data mapping a plurality of group types to a plurality of regularization types and based on group types of the plurality of groups, and generate at least one regularization term including a coefficient for the variables included in the groups for the at least one regularization type;
selecting an evaluation function to be used among a plurality of evaluation functions, depending on whether or not there is an overlap of variables among the plurality of groups, each evaluation function including the coefficients for the plurality of variables and the plurality of variables,
generating an objective function including the at least one regularization term generated for the at least one regularization type and including the selected evaluation function, and optimizing the objective function based on the operation data to estimate values of a plurality of coefficients for the plurality of variables;
selecting only variables with an absolute value greater than zero from the plurality of variables based on the values of the estimated coefficients;
identify, as drawing variables, the selected variables and their structurally adjacent variables determined based on the structure data;
generating, in a form of a table or graph, variable structure data that visually represents the drawing variables and structural relationships among the drawing variables, wherein the selected variables correspond to components having more influence on occurrence of the predetermined event among the plurality of components in the target apparatus and the non-selected variables correspond to components less influence on occurrence of the predetermined event among the plurality of components;
displaying the variable structure data visibly to a user;
constructing, based on a value of the coefficient estimated for the selected variable and the selected variable, a model of detecting occurrence of the predetermined event in the target apparatus;
receiving an operation data acquired by the plurality of sensors via a network;
calculating an output value of the model based on the operation data received, and detecting the occurrence of the predetermined event in the target apparatus based on the output value; and
displaying the variable structure data so that the user is able to specify a component that caused the occurrence of the predetermined event among the plurality of components in the target apparatus to one of the components corresponding to the selected variables, wherein
the operation data includes first operation data acquired in a first period and second operation data acquired in a second period later than the first period,
the processes comprises performing processing on the first operation data to select first variables being the variables with the absolute value greater than zero from the plurality of variables, and performing processing on the second operation data to select second variables being the variables with the absolute value greater than zero from the plurality of variables,
determining whether a first set of drawing corresponding to the first operation data structurally differs a second set of drawing variables corresponding to the second operation data and, when a structural difference is detected between the first and second sets of drawing variables, outputting an alert indicating that a structural mechanism of occurrence of the predetermined event in the target apparatus has changed.
14 . A non-transitory computer readable medium having a computer program stored therein which causes a computer to perform processes comprising:
reading operation data of a target apparatus including a plurality of variables obtained by sensing a plurality of components in the target apparatus with a plurality of sensors monitoring the plurality of components in the target apparatus and structure data representing structural relation among the plurality of variables from at least one memory, the operation data including whether or not a predetermined event has occurred;
grouping variables having a same type of structural relation among the plurality of variables in the operating data into a same group based on the structure data to generate a plurality of groups of at least one group type corresponding to a type of structural relation of variables in the groups, each of the plurality of variables being included in one of the plurality of groups;
determining at least one regularization type corresponding to the plurality of groups based on mapping data mapping a plurality of group types to a plurality of regularization types and based on group types of the plurality of groups, and generate at least one regularization term including a coefficient for the variables included in the groups for the at least one regularization type;
selecting an evaluation function to be used among a plurality of evaluation functions, depending on whether or not there is an overlap of variables among the plurality of groups, each evaluation function including the coefficients for the plurality of variables and the plurality of variables,
generating an objective function including the at least one regularization term generated for the at least one regularization type and including the selected evaluation function, and optimizing the objective function based on the operation data to estimate values of a plurality of coefficients for the plurality of variables;
selecting only variables with an absolute value greater than zero from the plurality of variables based on the values of the estimated coefficients;
identify, as drawing variables, the selected variables and their structurally adjacent variables determined based on the structure data;
generating, in a form of a table or graph, variable structure data that visually represents the drawing variables and structural relationships among the drawing variables components in the target apparatus, wherein the selected variables correspond to components having more influence on occurrence of the predetermined event among the plurality of components in the target apparatus and the non-selected variables correspond to components less influence on occurrence of the predetermined event among the plurality of components;
displaying the variable structure data visibly to a user;
constructing, based on a value of the coefficient estimated for the selected variable and the selected variable, a model of detecting occurrence of the predetermined event in the target apparatus;
receiving an operation data acquired by the plurality of sensors via a network;
calculating an output value of the model based on the operation data received, and detecting the occurrence of the predetermined event in the target apparatus based on the output value; and
displaying the variable structure data so that the user is able to specify a component that caused the occurrence of the predetermined event among the plurality of components in the target apparatus to one of the components corresponding to the selected variables, wherein
the operation data includes first operation data acquired in a first period and second operation data acquired in a second period later than the first period,
the processes comprises performing processing on the first operation data to select first variables being the variables with the absolute value greater than zero from the plurality of variables, and performing processing on the second operation data to select second variables being the variables with the absolute value greater than zero from the plurality of variables,
determining whether a first set of drawing corresponding to the first operation data structurally differs a second set of drawing variables corresponding to the second operation data and, when a structural difference is detected between the first and second sets of drawing variables outputting an alert indicating that a structural mechanism of occurrence of the predetermined event in the target apparatus has changed.